Cardiovascular phenotyping, AI, radiomics, and outcomes on planning CT.

Cardiovascular Phenotyping, Artificial Intelligence, Radiomics, and Clinical Outcomes Using Radiation Treatment Planning Computed Tomography in Patients Receiving Radiation Therapy: A Retrospective Multidisciplinary Cardiovascular Oncology Study.

Background

Cancer and cardiovascular disease continue to represent the two leading causes of morbidity and mortality throughout the world. Advances in oncology have dramatically improved survival across numerous malignancies, creating an expanding population of cancer survivors whose long-term health is increasingly influenced by cardiovascular disease. At the same time, radiation oncology has become progressively more sophisticated through image-guided treatment planning, intensity-modulated radiation therapy, adaptive radiation therapy, stereotactic body radiotherapy, proton therapy, and other advanced technologies. Nearly every patient undergoing radiation therapy receives a treatment planning computed tomography examination before treatment initiation. These planning CT examinations constitute one of the largest repositories of routinely acquired thoracic imaging in modern medicine, yet much of the cardiovascular information embedded within these studies remains unused during routine clinical care.

The proposed retrospective investigation seeks to transform radiation treatment planning CT imaging from a modality used primarily for oncologic targeting into a comprehensive platform for cardiovascular phenotyping, risk prediction, artificial intelligence development, radiomics discovery, and precision cardio-oncology. Rather than requiring additional imaging examinations, additional radiation exposure, or changes in patient care, this study will leverage existing clinical imaging obtained during routine cancer treatment planning. By integrating imaging features with radiation dosimetry, clinical characteristics, laboratory values, cardiovascular outcomes, oncologic outcomes, and longitudinal follow-up, the study aims to establish a highly scalable research platform capable of advancing cardiovascular oncology through secondary analysis of routinely acquired clinical data.

This retrospective study represents the foundational phase of a larger multidisciplinary cardiovascular oncology research program that will ultimately encompass prospective observational investigations, implementation science, machine learning, imaging biomarker validation, digital health integration, clinical decision support, and future multicenter collaborations.

Modern radiation treatment planning CT examinations contain substantially more information than is currently extracted for routine patient care. Although their primary purpose is target delineation and treatment planning, these examinations simultaneously visualize numerous cardiovascular structures, vascular beds, extracardiac tissues, musculoskeletal structures, pulmonary anatomy, adipose tissue distribution, skeletal integrity, and other biological features that may reflect both baseline cardiovascular health and future susceptibility to cardiovascular complications.

Existing evidence suggests that coronary artery calcification, thoracic aortic calcification, valvular calcification, cardiac chamber enlargement, epicardial adipose tissue, skeletal muscle mass, sarcopenia, visceral adiposity, pulmonary artery enlargement, emphysema, vertebral bone density, and numerous additional imaging-derived biomarkers associate with cardiovascular outcomes in diverse populations. However, relatively few investigations have comprehensively integrated these imaging biomarkers with individualized radiation dose distributions and subsequent cardiovascular events among patients undergoing radiation therapy.

Treatment-planning CT examinations therefore represent an extraordinary opportunity for opportunistic cardiovascular screening without requiring additional patient burden. Every existing planning CT may serve simultaneously as an oncologic planning study and a cardiovascular imaging examination. The retrospective analysis proposed herein seeks to quantify these imaging biomarkers, evaluate their relationships with cardiovascular outcomes, develop artificial intelligence algorithms capable of automated extraction, and establish reproducible methodologies for future prospective implementation.

The primary objective of this retrospective investigation is to develop and validate a comprehensive cardiovascular oncology imaging platform utilizing existing radiation treatment planning computed tomography examinations obtained during routine clinical care. The study will evaluate whether quantitative cardiovascular imaging biomarkers extracted from treatment planning CT examinations improve prediction of subsequent cardiovascular outcomes beyond conventional clinical risk assessment alone.

Primary Hypothesis

Baseline cardiovascular imaging biomarkers extracted from routine radiation treatment planning computed tomography examinations independently predict future cardiovascular morbidity and mortality among patients receiving radiation therapy.

Secondary Hypotheses

We further hypothesize that quantitative imaging biomarkers obtained from planning CT examinations provide complementary prognostic information beyond traditional cardiovascular risk factors and conventional oncologic characteristics. Specifically, coronary artery calcification burden, thoracic vascular calcification, epicardial adipose tissue volume, skeletal muscle composition, body composition metrics, cardiac chamber morphology, pulmonary vascular dimensions, vertebral bone density, and radiomic imaging signatures will each demonstrate independent associations with clinically meaningful cardiovascular outcomes.

We additionally hypothesize that radiation dose distributions delivered to individual cardiac substructures interact with baseline cardiovascular phenotype to influence subsequent development of cardiovascular toxicity. Consequently, patients with similar radiation dose exposure may experience markedly different cardiovascular outcomes depending upon their baseline cardiovascular imaging characteristics.

Finally, we hypothesize that artificial intelligence algorithms developed using treatment planning CT examinations will accurately and reproducibly automate extraction of cardiovascular imaging biomarkers, thereby enabling scalable implementation across diverse clinical environments.

The proposed study addresses one of the most important emerging challenges in contemporary medicine: the intersection between cardiovascular disease and cancer survivorship. Improvements in cancer therapy continue to increase long-term survival, shifting clinical attention toward treatment-related toxicities and competing causes of mortality. Cardiovascular disease has become one of the leading causes of death among long-term cancer survivors, highlighting the urgent need for improved cardiovascular risk assessment before, during, and after cancer therapy.

Despite this growing need, most treatment planning CT examinations are interpreted exclusively for radiation planning purposes, leaving substantial cardiovascular information unexplored. Every year, millions of planning CT examinations are obtained worldwide. Each examination represents a potential opportunity for opportunistic cardiovascular screening without additional imaging, additional radiation exposure, additional patient appointments, or increased healthcare costs.

The proposed investigation therefore has the potential to redefine the clinical utility of radiation treatment planning CT imaging. Rather than functioning solely as an oncologic planning tool, planning CT examinations may become a foundational component of precision cardiovascular oncology. This paradigm shift could substantially improve individualized cardiovascular risk assessment, optimize surveillance strategies, inform multidisciplinary treatment planning, facilitate earlier preventive interventions, and ultimately improve both cardiovascular and oncologic outcomes.

Furthermore, the methodologies developed through this retrospective investigation are intentionally designed to support future prospective implementation. Artificial intelligence algorithms generated through this work may ultimately provide automated cardiovascular phenotyping immediately after planning CT acquisition, integrating seamlessly into routine radiation oncology workflows without increasing physician workload.

The remarkable evolution of modern oncology has fundamentally transformed cancer from an immediately life-threatening diagnosis for many individuals into a chronic disease with progressively improving long-term survival. Across virtually every major malignancy, advances in surgical oncology, radiation oncology, systemic chemotherapy, targeted molecular therapies, immunotherapy, cellular therapies, and precision medicine have contributed to substantial improvements in survival over the past several decades. As the population of cancer survivors continues to expand, the clinical priorities of oncology have likewise evolved beyond achieving tumor control alone to encompass preservation of long-term health, optimization of quality of life, reduction of treatment-related toxicities, and prevention of competing causes of mortality.

Cardiovascular disease has emerged as one of the most important determinants of long-term outcomes among cancer survivors. Numerous epidemiologic investigations have demonstrated that cardiovascular disease frequently becomes the leading non-cancer cause of death following successful cancer treatment. In several malignancies, particularly breast cancer, lymphoma, thoracic malignancies, prostate cancer, and childhood cancers, cardiovascular mortality increasingly rivals or exceeds mortality from recurrent malignancy among long-term survivors. Consequently, cardiovascular oncology has developed into a rapidly expanding multidisciplinary subspecialty dedicated to understanding, preventing, identifying, monitoring, and treating cardiovascular disease throughout the continuum of cancer care.

Historically, much of cardio-oncology has focused upon systemic cancer therapies. Anthracyclines, HER2-targeted therapies, vascular endothelial growth factor inhibitors, tyrosine kinase inhibitors, proteasome inhibitors, immune checkpoint inhibitors, androgen deprivation therapy, fluoropyrimidines, platinum agents, and numerous additional therapies have each been associated with unique cardiovascular complications. Radiation therapy similarly contributes to cardiovascular injury through mechanisms involving endothelial dysfunction, accelerated atherosclerosis, chronic inflammation, fibrosis, microvascular injury, myocardial remodeling, valvular degeneration, conduction abnormalities, and pericardial disease. Importantly, radiation-associated cardiovascular injury frequently develops years or decades after treatment, emphasizing the importance of identifying patients at greatest risk before treatment begins.

Despite considerable advances in radiation treatment planning, relatively limited information currently guides individualized cardiovascular risk prediction before radiation therapy. Clinical risk assessment typically relies upon traditional cardiovascular risk factors, baseline medical history, and clinician judgment. Although these factors remain critically important, they likely capture only a portion of an individual's underlying cardiovascular vulnerability. Increasing evidence suggests that routinely acquired imaging examinations contain an extraordinary amount of latent biological information capable of substantially enhancing cardiovascular risk assessment.

Every patient receiving external beam radiation therapy undergoes treatment planning computed tomography before initiation of treatment. These planning CT examinations are indispensable for target delineation, treatment planning, dose optimization, organ-at-risk contouring, and image-guided treatment delivery. Worldwide, millions of treatment planning CT examinations are acquired annually. While these studies primarily serve radiation oncology, they simultaneously depict the heart, coronary arteries, thoracic aorta, pulmonary vasculature, cardiac chambers, mediastinum, lungs, skeletal musculature, adipose tissue, osseous structures, liver, and numerous additional anatomical structures with remarkable detail. Consequently, each planning CT examination constitutes a comprehensive physiologic snapshot obtained immediately before cancer treatment begins.

The overwhelming majority of this information currently remains unused. Radiation oncologists appropriately concentrate upon target definition and organ-at-risk protection. Diagnostic radiologists may not routinely interpret treatment planning CT examinations because image acquisition protocols differ from conventional diagnostic CT imaging. Cardiologists rarely review these examinations despite their rich cardiovascular content. As a result, numerous clinically meaningful cardiovascular imaging biomarkers frequently remain unrecognized despite already existing within routinely acquired clinical imaging.

Coronary artery calcification illustrates one of the most compelling examples of opportunistic cardiovascular imaging. Coronary artery calcification represents one of the strongest imaging predictors of future atherosclerotic cardiovascular disease, myocardial infarction, stroke, and cardiovascular mortality across multiple populations. Although coronary artery calcium scoring traditionally requires electrocardiographically gated cardiac CT examinations, numerous investigations have demonstrated that visually assessed or quantitatively estimated coronary calcification from non-gated thoracic CT examinations correlates strongly with dedicated calcium scoring examinations and predicts subsequent cardiovascular events. Treatment-planning CT examinations therefore represent an underutilized opportunity for identifying previously unrecognized coronary artery disease before initiation of potentially cardiotoxic cancer therapies.

Similarly, thoracic aortic calcification reflects cumulative vascular injury and systemic atherosclerosis. Aortic valve calcification predicts future valvular disease progression. Mitral annular calcification associates with atrial fibrillation, conduction abnormalities, stroke, and cardiovascular mortality. Epicardial adipose tissue functions as a metabolically active inflammatory organ influencing coronary artery disease and myocardial remodeling. Cardiac chamber enlargement reflects underlying structural heart disease. Pulmonary artery enlargement suggests pulmonary hypertension. Skeletal muscle quantity and quality provide objective measures of frailty, sarcopenia, nutritional reserve, and physiologic resilience. Vertebral bone density reflects osteoporosis risk while simultaneously serving as a marker of biological aging. Visceral adiposity and body composition influence inflammation, metabolic disease, cardiovascular risk, and cancer outcomes. Pulmonary parenchymal abnormalities identify chronic lung disease that may modify cardiopulmonary reserve.

Collectively, these imaging biomarkers provide a multidimensional characterization of baseline cardiovascular health that extends far beyond conventional clinical assessment. When integrated with clinical characteristics, laboratory values, medication histories, cardiovascular imaging studies, radiation dosimetry, cancer characteristics, treatment exposures, and longitudinal outcomes, these imaging features have the potential to substantially improve individualized cardiovascular risk prediction throughout the cancer continuum.

Recent developments in artificial intelligence further enhance the transformative potential of treatment planning CT imaging. Deep learning methodologies now permit rapid automated segmentation of cardiac structures, vascular anatomy, skeletal muscle, adipose tissue, pulmonary parenchyma, osseous structures, and additional organs with high reproducibility. Automated extraction of quantitative imaging biomarkers enables large-scale phenotyping across thousands of patients with minimal manual intervention. Radiomic analyses further expand this opportunity by quantifying imaging texture, heterogeneity, spatial organization, shape, intensity distributions, and higher-order mathematical features that may reflect underlying biological processes invisible to the human observer.

The convergence of routine clinical imaging, radiation dosimetry, artificial intelligence, radiomics, and longitudinal electronic health record data creates an unprecedented opportunity for precision cardiovascular oncology. Rather than obtaining additional imaging examinations, the proposed study capitalizes upon information that already exists within standard clinical workflows. This approach minimizes patient burden while maximizing scientific value, healthcare efficiency, and scalability across diverse healthcare systems.

The present retrospective investigation therefore represents the foundational scientific component of a long-term cardiovascular oncology imaging program jointly developed through Dartmouth Radiation Oncology, Heart Innovation and Equity Research (HIER) Institute, and Heart Spark Research and Innovation Institute. The retrospective study is intentionally designed to establish standardized imaging methodologies, validate quantitative imaging biomarkers, develop artificial intelligence algorithms, characterize relationships between cardiovascular phenotype and radiation exposure, and generate preliminary data supporting future prospective implementation, multicenter collaboration, investigator-initiated clinical trials, and externally funded translational research initiatives.

By transforming routine radiation treatment planning CT examinations into comprehensive cardiovascular phenotyping studies, this research seeks to redefine the role of imaging within cardiovascular oncology. Ultimately, every planning CT examination has the potential to become simultaneously a cancer treatment planning study, a cardiovascular screening examination, a precision medicine platform, an artificial intelligence training dataset, and a foundation for individualized survivorship care. The long-term vision extends beyond retrospective research alone and aspires to establish a learning health system in which continuously acquired clinical imaging informs increasingly accurate prediction, prevention, and personalization of cardiovascular care for patients with cancer.

The purpose of this retrospective investigation is to establish a comprehensive cardiovascular oncology imaging platform capable of extracting clinically meaningful cardiovascular biomarkers from existing radiation treatment planning computed tomography examinations acquired during routine cancer care. These biomarkers will be integrated with radiation treatment plans, dose distributions, electronic health record data, laboratory testing, cardiovascular imaging, oncologic characteristics, medications, and longitudinal clinical outcomes to better understand the biological relationships between baseline cardiovascular phenotype, cancer treatment, radiation exposure, and subsequent cardiovascular events.

Beyond evaluating individual imaging biomarkers, the study seeks to create an extensible research infrastructure supporting future artificial intelligence development, radiomics investigations, implementation science, digital health integration, precision survivorship strategies, multicenter collaborations, and prospective observational studies. The resulting platform is intended to facilitate continuous scientific discovery while accelerating translation of imaging-derived cardiovascular biomarkers into routine clinical care.

Overall Study Objective

The overarching objective of this retrospective investigation is to establish a comprehensive cardiovascular oncology imaging research platform that transforms routinely acquired radiation treatment planning computed tomography examinations into multidimensional cardiovascular phenotyping studies capable of informing individualized cardiovascular risk prediction before, during, and after cancer therapy. The study seeks to determine whether imaging-derived cardiovascular biomarkers extracted from existing planning CT examinations improve prediction of cardiovascular outcomes beyond traditional clinical risk factors while simultaneously creating a robust scientific infrastructure for future artificial intelligence development, radiomics discovery, implementation science, and prospective translational research.

The study further aims to characterize the complex biological interactions among preexisting cardiovascular disease, cancer characteristics, radiation treatment planning, radiation dose distributions, systemic therapies, demographic factors, body composition, vascular aging, inflammatory phenotypes, and long-term cardiovascular outcomes. Rather than evaluating isolated cardiovascular imaging markers individually, this investigation adopts an integrated systems biology framework in which multiple imaging biomarkers are considered collectively as components of an individual's overall cardiovascular phenotype.

An equally important objective is the establishment of a scalable imaging repository capable of supporting numerous future investigations without requiring repeated reconstruction of research infrastructure. Every treatment planning CT examination included within the retrospective cohort will become part of a continuously expanding cardiovascular oncology imaging resource supporting future hypothesis generation, biomarker validation, machine learning algorithm development, educational initiatives, multicenter collaborations, and investigator-initiated clinical trials.

Primary Objective

The primary objective is to determine whether quantitative cardiovascular imaging biomarkers extracted from routine radiation treatment planning computed tomography examinations independently predict subsequent major adverse cardiovascular events following cancer treatment.

Major adverse cardiovascular events will include myocardial infarction, ischemic stroke, transient ischemic attack, hospitalization for heart failure, new diagnosis of cardiomyopathy, clinically significant arrhythmias, atrial fibrillation or flutter requiring treatment, ventricular tachyarrhythmias, complete heart block requiring permanent pacemaker implantation, cardiovascular death, coronary revascularization, symptomatic peripheral arterial disease, pulmonary hypertension requiring treatment, clinically significant valvular heart disease, constrictive pericarditis, radiation-associated pericardial disease, myocarditis, cardiac transplantation, mechanical circulatory support, and all-cause mortality.

The study will evaluate both composite cardiovascular endpoints and individual cardiovascular outcomes separately to identify imaging biomarkers demonstrating disease-specific predictive value.

Secondary Objectives

A major secondary objective is to determine whether incorporation of quantitative cardiovascular imaging biomarkers into conventional cardiovascular risk assessment models significantly improves discrimination, calibration, and clinical decision-making compared with traditional clinical risk factors alone.

The study additionally seeks to characterize baseline cardiovascular disease burden across diverse malignancies, treatment strategies, age groups, racial and ethnic populations, biological sexes, socioeconomic strata, geographic regions, and cardiovascular risk profiles. This objective recognizes that baseline cardiovascular phenotype may differ substantially across cancer populations and may influence both cardiovascular toxicity and overall survival.

Another important objective involves quantification of radiation dose delivered to individual cardiac substructures and assessment of relationships between localized radiation exposure and subsequent cardiovascular outcomes. Particular attention will be devoted to dose distributions involving the left anterior descending coronary artery, left ventricle, right ventricle, left atrium, right atrium, ascending aorta, aortic valve, mitral valve, pulmonary arteries, pericardium, and cardiac conduction system whenever contouring methodologies permit accurate localization.

The study also seeks to determine whether baseline cardiovascular imaging phenotype modifies susceptibility to radiation-associated cardiovascular injury. Individuals with extensive coronary artery calcification, diffuse vascular disease, increased epicardial adipose tissue, or advanced biological aging may demonstrate differential vulnerability to radiation-induced cardiovascular injury despite similar radiation dose exposure.

Additional secondary objectives include evaluation of associations between imaging-derived frailty markers and treatment tolerance, relationships between body composition and overall survival, interactions between skeletal muscle quality and cardiovascular outcomes, and associations among pulmonary imaging biomarkers, cardiopulmonary reserve, and treatment-related complications.

Exploratory Objectives

The exploratory objectives of this investigation are intentionally expansive because the imaging repository established through this study is expected to serve as the scientific foundation for numerous future investigations.

One exploratory objective is the discovery of previously unidentified imaging biomarkers predictive of cardiovascular toxicity through comprehensive radiomic feature extraction. Rather than restricting analyses to predefined anatomical measurements, advanced radiomic methodologies will quantify thousands of quantitative imaging features describing tissue heterogeneity, spatial organization, texture, morphology, attenuation distributions, wavelet transformations, and higher-order mathematical representations of biological structure.

Another exploratory objective involves development of deep learning algorithms capable of fully automated cardiovascular phenotyping using treatment planning CT examinations. These algorithms will ultimately seek to identify coronary artery calcification, vascular calcification, valvular calcification, cardiac chamber enlargement, skeletal muscle characteristics, adipose tissue distribution, pulmonary vascular measurements, vertebral bone density, and additional biomarkers without manual intervention.

Further exploratory investigations may include digital twin development, integration of genomic information where available, multimodal prediction models incorporating laboratory biomarkers and imaging features, longitudinal assessment of cardiovascular aging following cancer therapy, unsupervised machine learning approaches for cardiovascular phenotyping, and development of individualized survivorship prediction models.

The retrospective nature of the present investigation permits flexible hypothesis generation while maintaining rigorous statistical methodology and careful control of false discovery rates during exploratory analyses.

Primary Endpoint

The primary endpoint of this investigation will be time to first major adverse cardiovascular event following initiation of radiation therapy. Time-to-event analyses will be measured beginning on the date of radiation treatment initiation and extending through the conclusion of available follow-up within the electronic health record or linked administrative databases.

Secondary Endpoints

Secondary endpoints include individual cardiovascular diagnoses, cardiovascular hospitalization, emergency department utilization for cardiovascular disease, progression of coronary artery disease, incident heart failure, reduction in left ventricular ejection fraction, clinically significant arrhythmias, permanent pacemaker implantation, implantable cardioverter-defibrillator implantation, coronary revascularization, cerebrovascular disease, venous thromboembolism, pulmonary hypertension, pericardial disease, myocarditis, valvular heart disease progression, cardiovascular mortality, cancer-specific mortality, progression-free survival, disease-free survival, treatment interruptions, treatment completion, treatment-related hospitalization, intensive care unit admission, and all-cause mortality.

Oncologic outcomes will also include local recurrence, regional recurrence, distant metastasis, second primary malignancy, overall response to treatment, and duration of remission where appropriate.

Imaging Endpoints

Quantitative imaging endpoints will include coronary artery calcium burden, thoracic aortic calcification burden, valvular calcification severity, epicardial adipose tissue volume, pericardial adipose tissue volume, visceral adiposity measurements, subcutaneous adiposity measurements, skeletal muscle area, skeletal muscle attenuation, sarcopenia indices, vertebral bone density, thoracic aortic diameter, pulmonary artery diameter, cardiac chamber volumes, cardiac silhouette measurements, pulmonary emphysema quantification, interstitial lung abnormalities, hepatic attenuation, body composition metrics, and comprehensive radiomic signatures extracted from cardiovascular structures. Whenever technically feasible, automated segmentation methodologies will generate three-dimensional volumetric measurements rather than relying solely upon linear dimensions or subjective visual grading.

Radiation Dosimetry Endpoints

Radiation dosimetry analyses will include mean heart dose, maximum heart dose, heart volume receiving predefined dose thresholds, biologically effective dose, equivalent dose in two-Gy fractions, and individualized dose distributions involving specific cardiac substructures. Dose-volume histogram parameters will be extracted for each contoured cardiovascular structure whenever contouring data are available. Three-dimensional dose maps will be co-registered with cardiovascular anatomical segmentations to permit detailed spatial analyses of localized radiation exposure.

Additional dosimetric analyses may evaluate cumulative dose heterogeneity, regional dose clustering, spatial overlap between calcified coronary arteries and radiation fields, and interactions between preexisting cardiovascular disease and radiation exposure patterns.

This is a single-institution retrospective observational cohort study conducted under an umbrella master IRB protocol. The study population consists of adult patients who received external beam radiation therapy at Dartmouth Radiation Oncology and for whom treatment planning computed tomography examinations, radiation treatment plans, and corresponding electronic health record data are available. All imaging, dosimetry, clinical, laboratory, medication, cardiovascular imaging, and outcome data will be obtained from existing clinical sources; no additional imaging, laboratory testing, or patient contact is required.

The umbrella architecture permits multiple predefined sub-studies to proceed under a single governance and regulatory framework, including retrospective feasibility studies of individual imaging biomarkers, radiation dose–substructure associations, artificial intelligence algorithm development and validation, radiomic signature discovery, disease-specific investigations (breast, lung, esophageal, lymphoma, head and neck, mediastinal tumors), and integrative multi-biomarker prognostic modeling. Future amendments will add additional sub-studies as the platform matures.

Inclusion Criteria

Adults aged 18 years or older who underwent treatment planning computed tomography at Dartmouth Radiation Oncology in association with external beam radiation therapy for any oncologic indication. Both curative and palliative treatment intents are eligible. All histologic diagnoses are eligible.

Exclusion Criteria

Patients whose planning CT examinations are unavailable, incomplete, or of insufficient technical quality to permit reliable cardiovascular phenotyping. Patients whose treatment plans, dose distributions, or electronic health record data are irretrievable will also be excluded from analyses that require those variables. Patients with documented objection to research use of their medical records will be excluded.

Sample Size

Because this investigation functions as an umbrella platform, sample size is defined at the sub-study level within statistical analysis plans. The overall cohort is expected to include several thousand patients treated at Dartmouth Radiation Oncology over the retrospective observation window, providing substantial statistical power for the majority of proposed analyses.

Eligible participants will consist of adult patients who underwent computed tomography simulation for external beam radiation therapy as part of routine clinical care within Dartmouth Radiation Oncology during the approved study period. Eligibility is intentionally designed to be broad because the overarching objective of this investigation is to characterize cardiovascular phenotype across the spectrum of contemporary oncology practice rather than within narrowly defined disease-specific populations. Broad eligibility will maximize external validity while supporting numerous future nested investigations under the umbrella protocol.

Participants must have undergone at least one radiation treatment planning computed tomography examination that is available within institutional imaging archives in Digital Imaging and Communications in Medicine (DICOM) format or another format permitting quantitative image analysis. Imaging quality must be sufficient for evaluation of cardiovascular structures appropriate for the planned analyses. Minor imaging artifacts will not automatically exclude examinations because many contemporary artificial intelligence methodologies are specifically designed to accommodate variable image quality encountered during routine clinical practice.

Patients receiving definitive radiation therapy, adjuvant radiation therapy, neoadjuvant radiation therapy, salvage radiation therapy, stereotactic body radiation therapy, stereotactic radiosurgery, palliative radiation therapy, proton therapy, adaptive radiation therapy, intensity-modulated radiation therapy, image-guided radiation therapy, volumetric modulated arc therapy, or additional clinically accepted radiation modalities will all be eligible. Inclusion of this broad spectrum of treatment strategies is expected to enhance understanding of cardiovascular phenotype across diverse therapeutic settings while supporting future treatment-specific subgroup analyses.

The study will include patients with newly diagnosed malignancies, recurrent malignancies, metastatic disease, multiple primary malignancies, hematologic malignancies, and solid tumors whenever radiation treatment planning CT examinations satisfy imaging eligibility criteria. Patients receiving concurrent systemic therapy, sequential systemic therapy, maintenance therapy, targeted therapy, immunotherapy, endocrine therapy, or combination treatment strategies likewise remain eligible because contemporary cancer care frequently involves multimodality treatment.

Participants may possess extensive preexisting cardiovascular disease or may have no recognized cardiovascular diagnoses before cancer treatment. Individuals with coronary artery disease, heart failure, atrial fibrillation, valvular heart disease, congenital heart disease, peripheral arterial disease, pulmonary hypertension, cerebrovascular disease, hypertension, dyslipidemia, diabetes mellitus, chronic kidney disease, obesity, tobacco exposure, inflammatory disorders, autoimmune disease, prior cardiovascular procedures, implanted cardiac devices, or prior cardiac surgery will remain eligible because these conditions represent clinically important components of baseline cardiovascular phenotype.

Patients with prior radiation exposure will also remain eligible because cumulative radiation exposure represents an important biological variable worthy of investigation. Rather than excluding these patients, previous radiation treatment history will be carefully documented and incorporated into statistical analyses whenever possible.

The treatment planning computed tomography examination must contain sufficient anatomical coverage to permit extraction of one or more predefined cardiovascular or systemic imaging biomarkers. Because radiation planning examinations vary substantially depending upon disease site, anatomical coverage requirements will differ according to the scientific objective under investigation.

Thoracic planning CT examinations generally provide visualization of the heart, coronary arteries, thoracic aorta, pulmonary arteries, mediastinum, lungs, thoracic musculature, vertebral bodies, ribs, and additional thoracic structures. These examinations will therefore support comprehensive cardiovascular phenotyping.

Breast cancer simulation CT examinations frequently provide outstanding visualization of the anterior heart, left anterior descending coronary artery region, ascending aorta, pulmonary vasculature, thoracic musculature, and chest wall. These examinations are expected to contribute substantially to investigations of radiation-associated coronary disease and cardiac substructure dosimetry.

Esophageal cancer simulation CT examinations typically encompass nearly the entire heart and mediastinum while exposing portions of the heart to substantial radiation doses. These examinations provide unique opportunities for studying dose-response relationships involving multiple cardiac substructures.

Lymphoma planning CT examinations similarly offer valuable visualization of mediastinal anatomy and frequently include younger patients with potentially prolonged survivorship, making these examinations particularly valuable for investigation of delayed cardiovascular toxicity.

Head and neck planning CT examinations may incompletely visualize the heart but often provide excellent visualization of carotid arteries, aortic arch, cervical vasculature, skeletal musculature, vertebral bodies, and body composition characteristics. Accordingly, these examinations remain valuable for selected analyses involving vascular aging, frailty, sarcopenia, and systemic cardiovascular risk.

Abdominal and pelvic planning CT examinations may contribute body composition analyses, visceral adiposity measurements, skeletal muscle quantification, hepatic attenuation, abdominal vascular calcification, lumbar vertebral bone density, and additional systemic biomarkers reflecting biological aging and cardiovascular health.

Consequently, anatomical coverage requirements will be determined individually for each predefined imaging endpoint rather than applying rigid universal inclusion criteria.

Patients younger than eighteen years of age at the time of radiation treatment planning will initially be excluded from the present retrospective investigation because pediatric cardiovascular oncology involves distinct biological mechanisms, developmental physiology, survivorship considerations, regulatory requirements, and outcome measures. Pediatric investigations may subsequently be developed under independent protocols specifically designed for childhood cancer survivors.

Planning CT examinations with severe technical deficiencies preventing reliable quantitative analysis will likewise be excluded from imaging analyses requiring accurate anatomical segmentation. Examples include incomplete image reconstruction, extensive corruption of DICOM files, catastrophic motion artifact, absent image metadata necessary for quantitative processing, or incomplete image acquisition that precludes evaluation of required anatomical structures.

However, image quality standards will intentionally remain permissive because contemporary clinical practice routinely encounters variability in acquisition techniques, patient positioning, respiratory motion, metallic hardware, implanted devices, and body habitus. Development of robust analytical methodologies capable of functioning under realistic clinical conditions represents an important scientific objective rather than a limitation.

Patients who specifically prohibit research use of their clinical information according to institutional policy, applicable regulations, or legal requirements will be excluded whenever such restrictions apply.

Beyond these limited exclusions, the study intentionally minimizes exclusion criteria in order to preserve representativeness, reduce selection bias, maximize statistical power, and create a repository reflective of real-world oncology populations.

Eligible participants will be identified through systematic interrogation of institutional radiation oncology scheduling systems, treatment planning databases, oncology information systems, electronic health records, radiology archives, treatment delivery records, and associated clinical databases.

Case ascertainment will proceed in several sequential stages. Initial electronic queries will identify patients who underwent radiation treatment planning computed tomography during the approved study interval. These preliminary records will then undergo automated linkage with electronic health records using secure institutional identifiers available only within approved institutional environments.

Following automated linkage, trained research personnel will perform standardized quality assurance procedures verifying imaging availability, treatment characteristics, demographic information, malignancy classification, radiation treatment dates, and availability of longitudinal follow-up.

Whenever uncertainty exists regarding eligibility, two independent investigators will review source documentation. Discrepancies will be resolved through consensus review involving senior investigators. This dual-review methodology is expected to improve consistency while reducing classification error.

All eligibility determinations will be documented within standardized electronic screening logs maintained under secure institutional governance.

Variable Collection Strategy

A defining characteristic of this investigation is the extraordinarily comprehensive scope of variable collection. Rather than limiting analyses to traditional cardiovascular risk factors alone, the study seeks to characterize the complete biological context within which cardiovascular toxicity develops.

Accordingly, data collection will encompass demographic characteristics, socioeconomic indicators where available, cancer diagnosis, histopathology, molecular subtype, staging information, treatment intent, radiation treatment characteristics, systemic therapy exposures, laboratory values, medications, cardiovascular history, imaging biomarkers, radiation dosimetry, body composition, radiomic features, artificial intelligence outputs, longitudinal outcomes, hospitalization history, procedural history, mortality, and survivorship measures.

Each variable included within the research database will possess a standardized operational definition documented within the study data dictionary. Variable definitions will undergo multidisciplinary review involving cardiologists, radiation oncologists, medical oncologists, radiologists, medical physicists, epidemiologists, statisticians, biomedical informaticians, and data scientists before final implementation.

Continuous variables will preferentially be recorded using their original numerical values rather than arbitrary categorical thresholds whenever scientifically appropriate. Preservation of continuous measurements enhances statistical flexibility, facilitates machine learning analyses, improves calibration of prediction models, and minimizes information loss. Categorical variables will similarly utilize standardized ontologies whenever possible to maximize interoperability with future multicenter collaborations.

Cardiovascular Variable Definitions

Baseline cardiovascular history will include comprehensive characterization of preexisting cardiovascular disease, cardiovascular procedures, implanted cardiac devices, cardiovascular medications, cardiovascular imaging findings, laboratory biomarkers, functional status, exercise capacity, smoking history, alcohol consumption, body mass index, blood pressure, diabetes status, renal function, lipid disorders, inflammatory diseases, sleep disorders, family history of premature cardiovascular disease, and additional clinically relevant factors.

Rather than relying exclusively upon diagnostic coding, cardiovascular diagnoses will preferentially be confirmed through physician documentation, imaging reports, procedural reports, laboratory findings, medication histories, and longitudinal clinical records whenever feasible. This multimodal verification strategy is expected to improve diagnostic accuracy while minimizing outcome misclassification.

Data Collection Procedures

Data abstraction will be performed by trained research personnel using standardized operating procedures developed specifically for this investigation. Structured electronic abstraction instruments will promote consistency across reviewers while facilitating quality assurance throughout repository development.

Each abstracted variable will undergo predefined validation procedures appropriate to its scientific importance. Random sampling, duplicate abstraction, interobserver agreement analyses, automated range checking, logic consistency algorithms, and periodic investigator review meetings will provide ongoing oversight of data quality. The study database will be continuously audited throughout development, ensuring that repository quality improves progressively as additional participants, imaging examinations, and longitudinal outcomes become available.

By combining rigorous clinical data abstraction with advanced quantitative imaging analysis, the resulting repository is expected to become one of the most comprehensive cardiovascular oncology imaging resources assembled from routine radiation treatment planning CT examinations, providing an enduring foundation for future scientific discovery, precision medicine, artificial intelligence development, and translational cardiovascular oncology research.

Imaging Acquisition & Repository

Treatment-planning CT examinations, corresponding radiation treatment plans, dose distributions, structure sets, cone-beam CT imaging (where available), diagnostic CT imaging, positron emission tomography, cardiac magnetic resonance imaging, echocardiography, and electrocardiograms will be retrieved from clinical imaging and treatment planning systems. Images will be de-identified according to institutional standards and stored in a secure research imaging repository under HIER Institute governance.

Radiation Dosimetry

Radiation dose distributions will be co-registered with cardiac substructure segmentations to derive per-structure dose–volume histograms for the whole heart, left and right ventricles, left and right atria, left anterior descending, right coronary and circumflex arteries, aortic, mitral, pulmonary and tricuspid valves, pericardium, conduction system, ascending and descending aorta, pulmonary arteries, esophagus, and great vessels.

Image Analysis

Cardiovascular imaging biomarkers will be extracted using validated quantitative imaging pipelines combining automated segmentation, densitometric analysis, morphometric measurement, and radiomic feature extraction. Independent human review will be performed on stratified random samples for quality assurance. Inter-rater reliability, intra-rater reliability, and algorithm–human agreement will be reported for each biomarker.

AI & Radiomics

Artificial intelligence algorithms will be trained and validated using standardized partitioning strategies, including patient-level train/validation/test splits and cross-validation. Radiomic features will be extracted using open, reproducible libraries. Feature harmonization (for example, ComBat-style methods) will be applied across scanners and protocols. Model performance will be evaluated using discrimination, calibration, decision-curve analysis, and clinical utility metrics appropriate to each outcome.

Outcomes Ascertainment

Cardiovascular and oncologic outcomes will be ascertained from the electronic health record, institutional cancer registries, cardiovascular registries, and mortality data sources. Standard definitions will be adopted for major adverse cardiovascular events, heart failure hospitalization, atrial fibrillation, myocardial infarction, stroke, valvular disease progression, pericardial disease, myocarditis, conduction disease, and cause-specific and all-cause mortality.

Imaging Philosophy

The imaging component of this investigation serves as the scientific foundation upon which the entire cardiovascular oncology research platform is constructed. Unlike conventional retrospective imaging studies that extract a limited number of predefined measurements, the present investigation seeks to comprehensively characterize cardiovascular anatomy, vascular biology, myocardial structure, extracardiac physiology, body composition, biological aging, and radiographic tissue characteristics using treatment planning computed tomography examinations acquired during routine radiation oncology care.

The overarching philosophy underlying this imaging program recognizes that every planning computed tomography examination contains substantially more biological information than is routinely used during radiation treatment planning. Rather than viewing these studies solely as geometric maps for radiation delivery, the proposed research conceptualizes each examination as a multidimensional biological dataset capable of providing quantitative information regarding cardiovascular health, systemic physiology, metabolic reserve, frailty, inflammation, vascular aging, and susceptibility to future cardiovascular injury.

Every imaging examination included in the repository will therefore undergo standardized quantitative analysis using reproducible methodologies developed through multidisciplinary collaboration among cardiovascular imaging specialists, radiation oncologists, medical physicists, radiologists, biomedical engineers, computer scientists, artificial intelligence investigators, statisticians, and clinical informaticians. These methodologies are intentionally designed to remain adaptable as imaging science continues to evolve, allowing future computational advances to be applied retrospectively to previously acquired examinations while preserving methodological consistency.

Imaging Acquisition

All imaging studies included within this investigation will have been obtained during routine clinical care before initiation of radiation therapy. No imaging examinations will be acquired specifically for research purposes. Consequently, participation in this retrospective investigation will not alter patient care, increase radiation exposure, modify imaging protocols, or introduce additional clinical procedures.

Planning computed tomography examinations will typically be acquired using multidetector computed tomography scanners employed routinely within the Department of Radiation Oncology. Scanner manufacturers, detector configurations, reconstruction algorithms, acquisition parameters, tube voltage, tube current modulation strategies, slice thickness, reconstruction kernels, respiratory instructions, patient positioning, intravenous contrast administration, and immobilization devices will vary according to standard institutional clinical practice over time.

Rather than attempting to eliminate this natural heterogeneity, the study intentionally embraces it. Modern artificial intelligence algorithms and quantitative imaging methodologies intended for clinical implementation must ultimately perform reliably across diverse scanners, evolving acquisition protocols, and realistic clinical environments. Accordingly, methodological robustness across heterogeneous imaging conditions constitutes an explicit scientific objective rather than a methodological limitation.

Comprehensive acquisition metadata will be extracted directly from Digital Imaging and Communications in Medicine headers whenever available. These metadata will include scanner manufacturer, scanner model, acquisition date, reconstruction parameters, voxel dimensions, matrix size, field of view, slice interval, gantry orientation, reconstruction kernel, patient positioning information, contrast administration details, respiratory instructions, acquisition timing, and additional technical parameters relevant to quantitative image analysis. Maintenance of complete acquisition metadata will facilitate subsequent harmonization procedures, sensitivity analyses, artificial intelligence development, and multicenter external validation.

Digital Imaging Repository

All treatment planning computed tomography examinations will be stored within a centralized secure imaging repository designed specifically for cardiovascular oncology research. Original Digital Imaging and Communications in Medicine datasets will remain preserved in their native format throughout the duration of the project. Derived datasets generated during preprocessing, segmentation, radiomic feature extraction, artificial intelligence analyses, and quantitative measurements will be stored independently from the original source images to preserve complete data provenance.

The repository will maintain permanent version control documenting every computational operation applied to each imaging examination. This includes preprocessing algorithms, segmentation software versions, artificial intelligence model versions, radiomic extraction software versions, quality assurance outcomes, investigator review dates, and subsequent revisions. Such documentation ensures complete reproducibility while supporting future regulatory review, scientific publication, multicenter collaboration, and external validation.

Every image entering the repository will receive a unique research identifier independent of clinical medical record numbers. The linkage between research identifiers and institutional identifiers will remain securely maintained within approved institutional environments accessible only to authorized study personnel according to approved regulatory procedures.

Image Preprocessing

Before quantitative analysis, each imaging examination will undergo standardized preprocessing designed to maximize reproducibility while preserving biologically meaningful imaging characteristics.

Image orientation will first be verified to ensure standardized anatomical alignment. Image integrity assessments will confirm completeness of image reconstruction, consistency of Digital Imaging and Communications in Medicine metadata, absence of corruption, and successful import into the research environment.

Voxel dimensions will then be evaluated. Because treatment planning computed tomography examinations frequently vary in slice thickness and in-plane spatial resolution, isotropic resampling procedures may be performed when appropriate for quantitative image analysis or artificial intelligence development. Resampling methodologies will utilize interpolation techniques selected according to current best practices within quantitative imaging research.

Intensity normalization procedures may subsequently be performed to reduce scanner-related variability while preserving tissue-specific attenuation characteristics measured in Hounsfield Units. Normalization strategies will undergo extensive validation to ensure preservation of clinically meaningful quantitative information.

Additional preprocessing operations may include artifact reduction, respiratory alignment, image denoising, intensity clipping, coordinate normalization, anatomical cropping, histogram standardization, image registration, and quality assurance assessments. Importantly, every preprocessing operation will be documented automatically within repository metadata, allowing complete reconstruction of every analytical pipeline.

Image Quality Assessment

Each examination will undergo comprehensive quality assessment before quantitative analysis. Image quality evaluation will include assessment of overall image completeness, respiratory motion artifact, metallic artifact, beam hardening, truncation artifact, patient positioning, anatomical coverage, reconstruction quality, attenuation consistency, and technical adequacy for predefined imaging objectives.

Rather than assigning simple binary classifications of acceptable or unacceptable quality, image quality will be characterized using standardized multidimensional scoring systems documenting strengths and limitations of each examination. Examinations unsuitable for one specific imaging analysis may remain fully suitable for numerous additional analyses. For example, extensive cardiac motion artifact may limit coronary artery calcium quantification while preserving body composition analysis, skeletal muscle measurements, pulmonary artery dimensions, vertebral bone density assessment, and radiomic characterization. Consequently, image quality determinations will be endpoint-specific rather than universally exclusionary.

Accurate segmentation of cardiovascular anatomy represents one of the most important methodological components of the entire investigation.

Initially, expert manual segmentation will establish reference standards for artificial intelligence development and algorithm validation. Experienced cardiovascular imaging investigators working collaboratively with radiation oncologists and radiologists will delineate predefined cardiovascular structures according to standardized contouring atlases developed specifically for this project.

The segmentation protocol will include the whole heart, left ventricle, right ventricle, left atrium, right atrium, interventricular septum, left ventricular myocardium, right ventricular myocardium, pericardium, ascending aorta, aortic arch, descending thoracic aorta, pulmonary trunk, right pulmonary artery, left pulmonary artery, superior vena cava, inferior vena cava when visualized, coronary sinus, aortic valve, mitral valve, tricuspid valve, pulmonary valve, and additional cardiovascular structures visible within individual examinations.

Whenever technically feasible, individual coronary arteries — including the left main coronary artery, left anterior descending coronary artery, left circumflex coronary artery, and right coronary artery — will undergo dedicated segmentation. Smaller coronary branches may also be segmented during selected methodological investigations evaluating high-resolution artificial intelligence approaches.

Beyond cardiovascular anatomy, segmentation protocols will encompass thoracic skeletal musculature, paraspinal muscles, pectoralis musculature, intercostal musculature, subcutaneous adipose tissue, visceral adipose tissue, epicardial adipose tissue, pericardial adipose tissue, lungs, vertebral bodies, liver when visualized, spleen when visualized, kidneys when visualized, and additional extracardiac anatomical structures supporting systemic phenotyping.

Segmentation quality will undergo rigorous quality assurance including independent review by experienced investigators, interobserver reproducibility analyses, intraclass correlation coefficients, Dice similarity coefficients, Hausdorff distance measurements, Bland–Altman analyses, and periodic calibration sessions designed to maintain contouring consistency throughout the study.

Although expert manual segmentation provides the highest anatomical fidelity, it is not scalable for repositories ultimately containing tens of thousands of imaging examinations. Accordingly, one of the principal methodological objectives of this investigation is development of automated segmentation algorithms capable of reproducing expert contours rapidly and accurately.

Deep convolutional neural networks, transformer-based architectures, hybrid segmentation frameworks, and emerging foundation imaging models will be evaluated throughout repository development. Automated segmentations will initially undergo expert review until predefined performance thresholds are consistently achieved.

Artificial intelligence algorithms will not replace human oversight during early implementation. Instead, automated segmentation will function as an efficiency-enhancing tool whose outputs remain subject to expert verification. As algorithm performance improves through iterative training, the proportion of manually reviewed examinations may gradually decrease according to predefined quality assurance metrics approved by the study leadership committee.

Eventually, successful artificial intelligence implementation is expected to reduce segmentation time from several hours per examination to only minutes while maintaining anatomical accuracy suitable for quantitative cardiovascular phenotyping.

Following segmentation, every imaging examination will undergo comprehensive quantitative cardiovascular phenotyping.

Rather than producing isolated measurements, the analytical framework conceptualizes each patient as possessing an integrated cardiovascular phenotype composed of numerous interrelated structural, vascular, metabolic, inflammatory, and body composition characteristics. These phenotypes will subsequently be analyzed both individually and collectively using traditional statistical methods, machine learning approaches, clustering algorithms, latent phenotype analyses, and systems biology methodologies.

The cardiovascular phenotype generated for each participant will ultimately consist of several hundred standardized imaging biomarkers extracted reproducibly from every eligible planning computed tomography examination. These biomarkers will form the quantitative foundation for all subsequent radiation dosimetry analyses, artificial intelligence development, radiomics investigations, longitudinal outcome prediction, and prospective validation studies.

Primary cardiovascular outcomes include major adverse cardiovascular events (a composite of cardiovascular death, non-fatal myocardial infarction, non-fatal stroke, and heart failure hospitalization), individual components of the composite, atrial fibrillation, ventricular arrhythmias, conduction disease requiring permanent pacemaker implantation, valvular heart disease, pericardial disease, myocarditis, pulmonary hypertension, cardiovascular hospitalization, cardiovascular mortality, and all-cause mortality. Secondary oncologic outcomes include overall survival, progression-free survival, cancer recurrence, and treatment-related complications. Exploratory outcomes include hospitalization, functional decline, and healthcare utilization.

Baseline characteristics will be summarized with descriptive statistics stratified by disease-specific cohorts. Associations between imaging biomarkers and time-to-event outcomes will be evaluated with Cox proportional hazards regression, competing-risks regression (Fine-Gray) where clinically appropriate, and Kaplan–Meier estimation. Multivariable models will adjust for age, sex, race, ethnicity, cardiovascular risk factors, cancer stage, cancer treatment (radiation dose, chemotherapy, immunotherapy, endocrine therapy), and relevant comorbidities.

Interaction terms will evaluate whether baseline cardiovascular phenotype modifies the effect of radiation dose on outcomes. Prognostic performance of imaging-based models will be quantified with Harrell's c-index, time-dependent AUC, calibration plots, and decision-curve analysis. Missing data will be handled with multiple imputation under missing-at-random assumptions where appropriate, with sensitivity analyses under alternative missingness assumptions. Multiple testing will be addressed through prespecified primary hypotheses and false-discovery-rate control for exploratory analyses.

For artificial intelligence and radiomics sub-studies, model development will follow TRIPOD-AI, CLAIM, and applicable reporting guidelines. External validation will be pursued through multicenter collaborations. Model performance drift will be monitored across time and scanner platforms.

All data will be stored on institutionally approved secure servers with encryption at rest and in transit, role-based access controls, comprehensive audit logging, and periodic security reviews. Direct identifiers will be replaced by study-specific identifiers, and the linking key will be maintained separately under restricted access. Data transfers between institutions will occur only under executed data-use agreements and only in de-identified or limited-dataset form as approved by the IRB. REDCap will serve as the primary structured data capture platform, with data dictionaries, case report forms, and quality-control procedures maintained by the Data Coordinating Center.

Purpose

A fundamental objective of the present investigation is to maximize efficiency, reproducibility, scalability, and sustainability through preferential utilization of structured clinical data already maintained within institutional information systems. Before initiation of large-scale data abstraction, the investigative team will perform a comprehensive institutional data inventory to identify the most complete and readily accessible source for every study variable.

Rather than relying upon manual review of individual electronic health records whenever possible, structured databases will serve as the preferred source of information because they improve reproducibility, reduce abstraction bias, decrease personnel requirements, facilitate automation, and support future expansion into multicenter collaborations. Accordingly, each study variable will be categorized according to anticipated availability within existing institutional data systems.

Before beginning data extraction, the investigative team will convene structured working sessions with representatives from every institutional group whose systems or expertise are required to construct the retrospective cohort. Confirming data availability, ownership, extraction pathways, and completeness in advance is expected to reduce manual abstraction by 60–90%, accelerate cohort assembly, and create a scalable foundation for future prospective studies and multicenter expansion.

Stakeholders to Convene

• Radiation Oncology
• Enterprise Data Warehouse
• Biomedical Informatics
• Epic Clarity administrators
• PACS administrators
• Radiation Physics
• RayStation / Eclipse administrators
• Cancer Registry
• Cardiology Informatics
• Clinical Research Office

Data Availability Matrix

For every study variable, the team will document the following attributes and maintain a living Data Availability Matrix updated as new institutional data sources become available.

Required Milestone

Completion of the Data Availability Assessment will be a required study start-up milestone and a formal deliverable of the Data Coordinating Center. The Matrix will be revisited at least annually thereafter and whenever a new institutional data source, ontology mapping, or extraction pathway becomes available. In large imaging repositories, this exercise consistently reduces manual abstraction by 60–90%, informs staffing plans, and materially improves the reproducibility and external validity of downstream analyses.

Following preprocessing, quality assurance, and anatomical segmentation, every eligible treatment planning CT will undergo comprehensive quantitative cardiovascular phenotyping using standardized methodologies developed specifically for this investigation. The objective is to convert routine clinical imaging into reproducible biological measurements that accurately characterize baseline cardiovascular health before initiation of radiation therapy.

Unlike traditional observational studies that measure only one or two cardiovascular variables, the present investigation characterizes each participant using hundreds to thousands of quantitative imaging biomarkers spanning vascular health, myocardial structure, adipose tissue biology, skeletal integrity, body composition, pulmonary physiology, and biological aging. Collectively, these measurements define an individualized cardiovascular phenotype and provide the foundation for subsequent statistical analyses, AI development, radiation dosimetry investigations, and prospective validation studies.

All quantitative analyses will be performed using standardized software pipelines operating under version-controlled computational environments. Every measurement is accompanied by metadata documenting software version, algorithm version, processing date, preprocessing parameters, investigator oversight, and quality assurance outcomes to maximize reproducibility and facilitate future multicenter harmonization and regulatory review.

Coronary Artery Calcification

Coronary artery calcification (CAC) is a principal cardiovascular imaging biomarker because of its well-established associations with future atherosclerotic cardiovascular disease, myocardial infarction, stroke, cardiovascular mortality, and overall mortality. Although planning CT is not routinely ECG-gated, visually assessed and quantitatively estimated CAC on non-gated thoracic CT correlates strongly with dedicated calcium scoring, enabling opportunistic identification of previously unrecognized coronary disease without additional imaging or radiation.

The left main, left anterior descending, left circumflex, and right coronary arteries will each undergo individual assessment when image quality permits. CAC burden will be characterized using visual ordinal grading, semiquantitative regional scoring, volumetric calcium measurements, calcium density metrics, automated ML-derived scores, and, when technically feasible, Agatston-equivalent methodologies adapted for non-gated imaging. The LAD receives particular attention because of its frequent proximity to radiation treatment fields in left-sided breast, mediastinal, esophageal, and other thoracic malignancies.

Thoracic Vascular Calcification

The ascending aorta, aortic arch, descending thoracic aorta, brachiocephalic artery, carotid origins when visualized, subclavian arteries, pulmonary arteries, and additional thoracic vessels will undergo systematic evaluation for calcification burden. Quantitative measurements include total calcification volume, regional distribution, circumferential involvement, plaque density, calcification morphology, and radiomic characterization of vascular mineralization. Three-dimensional vascular reconstructions will permit evaluation of calcification topology and regional clustering when technically feasible.

Cardiac Valve Calcification

Calcification of the aortic valve, mitral annulus, mitral leaflets, tricuspid valve, and pulmonary valve will be systematically characterized. Aortic valve calcification is associated with progressive aortic stenosis and cardiovascular mortality; mitral annular calcification is associated with atrial fibrillation, conduction disease, stroke, heart failure, and mortality. Analyses include volumetric burden, regional distribution, leaflet involvement, annular involvement, radiomic texture, and machine learning–derived phenotypic classifications, with longitudinal follow-up characterizing progression to clinically significant valvular disease.

Cardiac Chamber Morphometry

Detailed quantitative morphometry will be performed for every examination with adequate anatomical visualization: left and right ventricular dimensions, atrial size, wall and septal thickness, cardiac silhouette, ventricular geometry, atrial remodeling, and overall morphology. Volumetric measurements are preferred over linear metrics for reproducibility. Predefined phenotypes associated with hypertension, cardiomyopathy, pulmonary hypertension, congenital disease, infiltrative cardiomyopathies, and prior myocardial infarction will be documented and integrated with echocardiography, cardiac MRI, biomarkers, and longitudinal outcomes.

Epicardial & Pericardial Adipose Tissue

Epicardial adipose tissue lies directly adjacent to myocardium and coronary arteries without fascial separation, permitting paracrine interactions that influence vascular inflammation, endothelial dysfunction, plaque instability, myocardial fibrosis, arrhythmogenesis, and adverse remodeling. Comprehensive quantification will include volume, regional distribution, attenuation, radiomic texture, spatial relationships with coronary arteries, and longitudinal outcome associations. Pericardial adipose tissue is analyzed separately because its biology differs from true epicardial fat despite frequent conflation in prior literature. Both compartments are treated as continuous variables to maximize statistical flexibility.

Body Composition

Planning CT uniquely enables whole-body physiologic characterization. Each examination will undergo body composition analysis including skeletal muscle area and attenuation, muscle radiomic features, visceral and subcutaneous adiposity, intramuscular adipose infiltration, hepatic attenuation, vertebral bone density, and thoracic/abdominal musculature when visualized. Both muscle quantity and quality (sarcopenia and myosteatosis) will be modeled as determinants of frailty, cardiovascular disease, chemotherapy tolerance, postoperative outcomes, hospitalization, and mortality.

Pulmonary Phenotyping

Pulmonary parenchyma will be systematically evaluated for emphysema, fibrosis, interstitial abnormalities, airway disease, pulmonary vascular remodeling, and radiation-induced lung injury. Pulmonary artery diameter, pulmonary artery–to–aorta ratio, vascular tortuosity, regional emphysema burden, fibrosis extent, and radiomic signatures will be quantified when technically feasible, improving prediction of cardiopulmonary complications during and after cancer treatment.

A defining strength of this program is comprehensive integration of quantitative cardiovascular phenotyping with individualized radiation dosimetry. Prior investigations have relied largely on summary measures such as mean or maximum heart dose, which incompletely characterize the heterogeneous spatial distribution of radiation across individual cardiovascular structures. This study constructs an anatomically resolved, three-dimensional cardiovascular dosimetric profile for every participant.

Radiation Treatment Planning Data

For every eligible participant, complete planning datasets will be obtained: planning CT examinations, RT Structure Sets, RT Plans, RT Dose objects, dose-volume histograms, fractionation schedules, beam arrangements, treatment techniques, IGRT records, adaptive replanning information, and treatment completion documentation. Original DICOM RT objects are preserved rather than relying on exported summaries, enabling future re-analysis while preserving spatial fidelity. Metadata describing planning software and algorithm versions, dose calculation methods, heterogeneity correction, optimization strategy, treatment machine, beam energy, MLC configuration, respiratory management, positioning, immobilization, and registration methodology will be retained when available.

Cardiac Substructure Dosimetry

Each segmented cardiovascular structure will be spatially registered to the 3D dose distribution using validated registration methodologies. Exposure metrics will be calculated separately for the whole heart, left and right ventricles, atria, interventricular septum, LAD, LCx, RCA, left main when feasible, ascending aorta, aortic arch, descending thoracic aorta, pulmonary trunk and branches, superior vena cava, pericardium, and each cardiac valve. Metrics include mean, median, maximum, minimum, percentiles, equivalent uniform dose, biologically effective dose, EQD2, dose heterogeneity and conformity indices, gradient measures, and volumetric dose parameters (V5–V50+).

Dose–Volume Analysis & Cumulative Exposure

Continuous dose distributions will be retained rather than collapsed into arbitrary categorical variables, enhancing statistical power and permitting identification of nonlinear dose–response relationships. Spatial analyses will evaluate dose gradients across myocardial regions, overlap between calcified coronary arteries and high-dose regions, and interactions among baseline disease burden, body composition, anatomical variability, and treatment geometry. When historical treatment records exist, cumulative lifetime thoracic radiation exposure will be estimated to characterize patients with prior radiation therapy — a subgroup at elevated risk for long-term remodeling and subsequent radiation-associated injury.

Image Registration

Rigid, deformable, landmark-based, and hybrid registration methodologies will be employed according to predefined analytical objectives. Registration accuracy is a critical determinant of precise substructure dosimetry and will undergo continuous quality assurance throughout repository development, with systematic validation before quantitative analyses proceed.

Radiomics mathematically characterizes tissue heterogeneity, spatial organization, intensity distributions, geometric complexity, and higher-order image features that often remain imperceptible during routine visual interpretation. Each segmented cardiovascular structure may generate thousands of quantitative features describing first-order statistics, histogram characteristics, gray-level co-occurrence and run-length and size-zone matrices, neighborhood gray-tone differences, wavelet decompositions, Laplacian transformations, fractal geometry, local binary patterns, and additional higher-dimensional representations of biological structure.

Rather than assuming a priori which imaging characteristics prove clinically important, radiomics permits unbiased discovery of previously unrecognized biomarkers associated with cardiovascular toxicity, overall survival, treatment response, and long-term survivorship. Feature selection will use rigorous dimensionality reduction, biological plausibility assessment, reproducibility analyses, stability testing across preprocessing methods, and independent validation whenever feasible. Harmonization techniques will address scanner variability while preserving biological signal. Radiomic signatures identified retrospectively may serve as the basis for future clinical decision-support tools capable of identifying patients at elevated cardiovascular risk immediately following routine radiation treatment planning.

Artificial intelligence is a central methodological pillar of this program. Its primary purpose is to make large-scale analysis of planning CT feasible, reproducible, and clinically scalable. Manual review of every cardiovascular structure, calcium region, muscle compartment, adipose depot, and dose relationship across a large institutional cohort would be prohibitive. AI transforms the imaging repository from a static archive into a dynamic discovery platform, focusing initially on automated segmentation, biomarker extraction, quality assessment, data harmonization, outcome prediction, and latent phenotype identification. All development occurs within secure approved research environments using de-identified or coded datasets whenever feasible.

Model Development Dataset

The dataset will be partitioned into training, validation, and internal test cohorts before final evaluation, stratified by cancer type, anatomic region, sex, age, scanner, contrast status, and treatment era when sample size permits. Training optimizes parameters; validation guides selection, hyperparameter tuning, calibration, and early stopping; the internal test cohort is held out for final assessment. Temporal validation trains on earlier treatment years and tests on later years to evaluate robustness across scanner technology, planning protocols, and practice patterns.

Segmentation Models

Targets include whole heart, chambers, myocardium, pericardium, thoracic aorta, pulmonary arteries, coronary artery regions, valves, lungs, skeletal muscle and adipose compartments, vertebral bodies, liver when visualized, and other structures relevant to cardiovascular oncology phenotyping. Ground-truth labels are generated through manual or semi-automated segmentation with senior expert review on a subset to establish reference-quality data. Performance is assessed with Dice similarity coefficients, Hausdorff distance, average surface distance, volume similarity, centerline distance for vascular structures, and clinically meaningful error thresholds. Robustness is evaluated across field of view, contrast enhancement, slice thickness, positioning, immobilization devices, and artifact burden — models must operate reliably on the imperfect, varied images of routine care.

Biomarker Extraction Models

After segmentation, AI models extract quantitative biomarkers including CAC, thoracic aortic calcium, valvular calcium, epicardial and pericardial adipose volume, skeletal muscle area and attenuation, visceral and subcutaneous adiposity, pulmonary artery diameter, aortic diameter, vertebral bone density, emphysema burden, and interstitial lung abnormalities. Calcification algorithms identify high-attenuation foci within anatomically constrained regions and distinguish coronary calcium from adjacent bone, contrast, surgical clips, radiation markers, and device leads. Body composition algorithms combine HU thresholds with anatomical segmentation and learned tissue classification, capturing muscle quality through attenuation, heterogeneity, and radiomic features to model frailty, sarcopenia, metabolic reserve, and biological aging.

Predictive Modeling

Predictive models will estimate risk of subsequent cardiovascular outcomes after radiation therapy, including MACE, heart failure, atrial fibrillation, MI, stroke, coronary revascularization, pacemaker implantation, pulmonary hypertension, pericardial disease, valvular disease, cardiovascular hospitalization, cardiovascular mortality, and all-cause mortality. Models incorporate clinical variables, imaging biomarkers, dosimetry, cancer characteristics, systemic therapy, laboratory values, medications, and longitudinal follow-up. Candidate methods include traditional and penalized regression, random forests, gradient boosting, survival forests, neural networks, deep survival models, and multimodal architectures — selected according to sample size, outcome frequency, interpretability, calibration, and scientific purpose. Performance is assessed by discrimination, calibration, decision-curve analysis, NRI, IDI, time-dependent AUC, Brier score, and clinically meaningful risk thresholds; calibration receives particular emphasis.

Explainable AI

Because cardiovascular oncology decisions require clinical trust, explainability is built in from the beginning. Methods include feature importance rankings, saliency maps, class activation maps, SHAP values, counterfactual explanations, uncertainty maps, and clinically interpretable risk factor summaries. Visual overlays support review of segmentation intent; predictive explainability identifies the biomarkers, clinical variables, dose characteristics, and treatment exposures contributing most to individual or group risk estimates. AI in this protocol is intended to serve as a lantern for scientific discovery — making hidden patterns visible while preserving human oversight.

Bias, Fairness & Generalizability

Models are evaluated across clinically meaningful subgroups — age, sex, race, ethnicity, cancer type, treatment site, body size, comorbidity burden, scanner type, contrast status, and treatment era — when sample size permits. The team assesses whether error differs systematically across groups and whether recalibration, retraining, or subgroup-specific evaluation is required before broader deployment. A single performance estimate is never assumed to represent all patients equally; analyses attend to representativeness, missingness, data quality, and structural differences in healthcare access that may influence observed outcomes.

Data Security for AI Workflows

All AI workflows occur within secure research computing environments approved by institutional information security and compliance offices. Identifiable clinical data remain within institutional systems unless a formal data use agreement, business associate agreement, reliance agreement, or other approved mechanism authorizes transfer. External collaborators receive only the minimum necessary data. Development datasets are de-identified, coded, or limited according to the approved IRB and HIPAA pathway. Access is role-based, with activity logs, file access records, version control, encryption, password protection, and secure storage. Derived models, segmentation outputs, radiomic features, and analytic datasets are stored with the protection appropriate to their re-identification risk.

The protocol supports a mobile-optimized research web application layer that allows authorized team members to manage study workflows efficiently across devices while preserving every existing page, form, dashboard, workflow, permission structure, and data element. The interface is responsive across desktop, tablet, and mobile screens and prioritizes study start-up, data availability mapping, imaging review queues, segmentation status, quality assurance flags, abstraction status, missing data reports, outcome adjudication tasks, manuscript tracking, meeting notes, regulatory documents, and budget monitoring.

On phones, complex tables convert into expandable cards. Forms support autosave, progressive disclosure, large touch targets, sticky navigation, accessible typography, and simplified review flows. Dashboards use responsive cards rather than dense grids so investigators can see study progress, pending reviews, protocol milestones, and data completeness at a glance.

A data source mapping module allows the Dartmouth team to record where each variable resides, whether it is structured, whether it can be extracted automatically, and whether manual review is needed — the operational companion to the Data Availability Assessment. The platform includes secure authentication, role-based access, audit logging, session timeout, encrypted storage, document version control, and compatibility with institutional security requirements, supporting research efficiency while respecting the regulatory boundaries of retrospective clinical data use.

Reliable identification of cardiovascular outcomes is a principal determinant of scientific validity in observational cardiovascular oncology research. The study uses a comprehensive strategy integrating structured EHR data, diagnostic coding, procedural records, imaging reports, laboratory testing, medication histories, hospitalization records, physician documentation, and mortality information. Structured sources serve as the initial scalable mechanism for identifying candidate events; identified events subsequently undergo validation per predefined adjudication procedures.

Primary Cardiovascular Outcome

The primary outcome is the first occurrence of a major adverse cardiovascular event following initiation of radiation therapy. MACE includes myocardial infarction, ischemic stroke, transient ischemic attack, hospitalization for heart failure, newly diagnosed cardiomyopathy, clinically significant ventricular arrhythmia, sustained atrial fibrillation requiring treatment, cardiovascular death, coronary revascularization, and additional clinically meaningful events specified in endpoint definitions. Time-to-event analyses measure elapsed time from radiation treatment initiation until the first qualifying event.

Secondary Cardiovascular Outcomes

Secondary outcomes analyzed independently include coronary artery disease progression, unstable angina, coronary revascularization procedures, HFrEF, HFpEF, myocarditis, pericarditis, constrictive pericarditis, pericardial effusion requiring intervention, pulmonary hypertension, valvular disease progression, permanent pacemaker and ICD and CRT implantation, ventricular assist device implantation, cardiac transplantation, venous thromboembolism, peripheral arterial disease, aortic aneurysm progression, aortic dissection, cardiovascular hospitalization, ED utilization for cardiovascular causes, and cardiovascular mortality. Each endpoint has a standardized operational definition documented within the study manual of procedures to ensure consistency across investigators and future multicenter collaborations.

Oncologic Outcomes

Cardiovascular health and cancer outcomes are deeply interconnected. The study collects local recurrence, regional recurrence, distant metastatic progression, second primary malignancies, disease-free survival, progression-free survival, overall survival, treatment interruptions, treatment completion, radiation-associated toxicities, hospitalization during treatment, ICU utilization, cancer-specific mortality, and all-cause mortality. Integrated cardiovascular–oncologic outcomes permit competing-risk analyses and support future survivorship prediction models.

Outcome Adjudication

Adjudication procedures are standardized before formal analyses. Candidate events identified through structured databases are confirmed using predefined diagnostic criteria. Myocardial infarction incorporates clinical documentation, cardiac biomarkers, ECG, coronary angiography, revascularization, and physician diagnosis. Heart failure incorporates hospitalization records, imaging, natriuretic peptides, physician documentation, medication changes, and echocardiographic evidence of ventricular dysfunction. Arrhythmia outcomes incorporate ECG, ambulatory monitoring, device interrogation, electrophysiology procedures, and cardiology documentation. Stroke incorporates neurological consultation, neuroimaging, discharge diagnosis, and procedural records. Residual uncertainty triggers additional manual review by investigators with cardiovascular expertise.

Mortality Ascertainment

Vital status is determined using institutional EHR data supplemented by additional approved sources when available. Date of death is recorded whenever reliably documented. Cause of death is classified when sufficient information exists; because attribution may be complex in advanced malignancy, multiple cause-of-death categories may be retained rather than binary classifications. Competing-risk methodologies are used when appropriate to distinguish cardiovascular from cancer-related mortality.

Longitudinal Event Surveillance

Participants contribute longitudinal information beginning with initiation of radiation therapy through the latest available follow-up within approved institutional records. Repeated cardiovascular imaging, laboratory testing, medication changes, outpatient visits, hospitalizations, ED encounters, cardiovascular procedures, recurrent malignancy, additional cancer therapies, and survivorship outcomes all contribute. Because radiation-associated cardiovascular injury frequently evolves over years to decades, prolonged follow-up substantially increases scientific value — supporting investigation of immediate treatment-related complications, intermediate remodeling, delayed toxicity, chronic vascular aging, and lifelong survivorship.

A multidisciplinary Endpoint Governance Committee oversees ongoing refinement of endpoint definitions, adjudication procedures, quality assurance, and methodological consistency throughout the project. Membership includes representatives from cardiovascular medicine, cardio-oncology, radiation oncology, medical oncology, radiology, medical physics, epidemiology, biostatistics, biomedical informatics, artificial intelligence, and clinical research operations.

Regular meetings review ambiguous cases, update operational definitions in response to evolving evidence, monitor adjudication consistency, and recommend protocol amendments when appropriate. This governance structure ensures that endpoint ascertainment remains rigorous, reproducible, transparent, and scientifically robust as the cardiovascular oncology imaging repository continues to expand.

This retrospective secondary use of existing clinical data meets the criteria for a waiver of informed consent and a waiver of HIPAA authorization under 45 CFR 46.116 and 45 CFR 164.512(i). The research involves no more than minimal risk to participants; the waiver will not adversely affect the rights and welfare of participants; the research could not practicably be carried out without the waiver (because prospectively contacting thousands of patients treated over many years would be prohibitive and could introduce substantial selection bias); the research could not practicably be carried out without access to protected health information; the privacy risks are minimized by strict de-identification, secure storage, limited access, and comprehensive data governance; and participants will be provided with pertinent information after participation only when clinically appropriate.

This study poses no more than minimal risk. Because all data are obtained from existing clinical sources, participants incur no additional imaging exposure, additional testing, additional appointments, or changes to their clinical care. The principal residual risk is the potential for loss of confidentiality, which is mitigated by extensive de-identification, secure storage, restricted access, encrypted transmission, audit logging, and staff training. Participants receive no direct individual benefit; however, the study is expected to generate substantial societal benefit by advancing cardiovascular oncology, informing risk prediction, and supporting future clinical trials that may improve care for future patients.

Ethical stewardship of these data are a central priority of the initiative. Governance is shared across HIER Institute, Heart Spark RII, and Dartmouth Radiation Oncology, with explicit commitments to equitable research design, community engagement, responsible artificial intelligence development, transparent authorship, and prospective publication planning.

The success of this investigation depends not only upon acquisition of high-quality imaging data but upon creation of a robust, secure, scalable, and reproducible research data ecosystem capable of supporting decades of cardiovascular oncology investigation. Accordingly, the protocol adopts a modern data management philosophy emphasizing interoperability, transparency, automation, reproducibility, longitudinal expansion, regulatory compliance, and scientific reuse.

Rather than constructing a database designed solely for the immediate retrospective analyses, this protocol establishes a permanent cardiovascular oncology research infrastructure capable of supporting future prospective studies, multicenter collaborations, artificial intelligence development, radiomics investigations, implementation science, clinical trials, educational initiatives, and translational precision medicine programs.

Every data element incorporated into the repository is considered part of a continuously evolving learning health system. The repository is intentionally designed to improve over time through iterative enhancement of data completeness, imaging analysis, computational methodology, artificial intelligence, and longitudinal clinical follow-up.

The research database consists of interconnected relational domains. Each participant receives a unique research identifier that serves as the primary linkage variable across all study domains, and each participant possesses a longitudinal master record linking every available imaging examination, treatment episode, laboratory measurement, cardiovascular event, oncologic outcome, and survival datum throughout the duration of available follow-up.

Principal domains include:

• Participant demographics
• Cancer diagnosis
• Oncologic treatment history
• Radiation treatment planning
• Treatment-planning computed tomography
• Radiation dosimetry
• Cardiovascular history
• Laboratory testing
• Medication history
• Cardiovascular imaging
• Hospitalizations and outpatient encounters
• Procedural history
• Longitudinal clinical outcomes
• Mortality information
• Quantitative imaging biomarkers
• Radiomic features
• Artificial intelligence outputs
• Quality assurance metrics
• Derived analytical datasets

The relational architecture permits independent updating of individual domains while preserving referential integrity across the entire repository. New imaging analyses, additional follow-up, emerging biomarkers, and future computational outputs can therefore be incorporated without redesign of the overall database structure.

The database integrates information from multiple institutional systems.

Primary clinical information originates from the institutional electronic health record. Radiation treatment information originates from the radiation oncology information system and treatment planning system. Imaging information originates from institutional Picture Archiving and Communication Systems and DICOM archives. Laboratory information originates from institutional laboratory information systems. Medication information originates from pharmacy databases and medication administration systems.

Cardiovascular diagnostic testing originates from cardiovascular imaging laboratories, electrocardiography systems, ambulatory monitoring databases, catheterization laboratories, and procedural reporting systems. Administrative information may originate from institutional enterprise data warehouses, billing systems, scheduling databases, mortality registries, and approved administrative datasets.

Future expansion may incorporate wearable devices, remote monitoring, patient-reported outcomes, biospecimen repositories, genomic resources, environmental exposure databases, geospatial information, and external cardiovascular registries under separate regulatory approvals.

Before initiation of formal data collection, the investigative team develops a comprehensive master data dictionary governing every variable included within the repository. Each variable possesses a standardized definition including variable name, display name, description, scientific rationale, source system, source table, source field, data type, permissible values, measurement units, collection frequency, missing value definitions, quality assurance procedures, transformation rules, derived variable dependencies, statistical classification, analytical priority, version history, and responsible investigator.

The master data dictionary functions as the authoritative reference document governing all data extraction, quality assurance, statistical analyses, artificial intelligence development, manuscript preparation, and future multicenter harmonization.

Clinical data originating from multiple institutional systems frequently exhibit differences in terminology, coding systems, units of measurement, documentation practices, and temporal resolution. All incoming data therefore undergo systematic harmonization before incorporation into the analytical repository.

Diagnostic terminology is standardized using internationally accepted ontologies whenever feasible. Medication names are normalized. Laboratory measurements undergo unit standardization. Radiation treatment variables utilize standardized nomenclature. Imaging-derived measurements follow predefined quantitative conventions. Temporal variables utilize consistent date and time formats. Missing data indicators are standardized. Categorical variables undergo consistent coding. Continuous variables retain original precision whenever scientifically appropriate.

This harmonization process substantially improves reproducibility while facilitating future integration with external institutions.

Every dataset entering the repository undergoes automated and manual validation. Automated validation evaluates missing values, impossible values, duplicate records, logical inconsistencies, date conflicts, range violations, internal contradictions, identifier mismatches, cross-table consistency, and unit consistency. Manual validation includes source document verification, duplicate abstraction, random sampling, investigator review, expert adjudication, interobserver agreement, and longitudinal consistency review. Quality assurance findings are documented prospectively throughout repository development.

Data quality is continuously monitored using predefined quantitative metrics including overall completeness, field-specific completeness, longitudinal completeness, imaging completeness, follow-up completeness, laboratory completeness, medication completeness, segmentation success rate, AI success rate, radiomics extraction success, quality assurance review completion, time from acquisition to processing, manual abstraction time, and automated extraction success. Each metric is reviewed periodically by the Data Governance Committee, and thresholds requiring corrective action are predefined before repository development begins.

Missing data represent an expected feature of retrospective observational investigations. Rather than attempting to eliminate missingness entirely, the study carefully characterizes missing data, understands the mechanisms responsible, and applies statistically appropriate analytical methodologies.

Patterns of missingness are evaluated descriptively. Variables are classified as missing completely at random, missing at random, missing not at random, structurally unavailable, institutionally unavailable, technically unavailable, or clinically unnecessary. Analytical strategies may include complete-case analyses, multiple imputation, inverse probability weighting, likelihood-based estimation, Bayesian approaches, machine learning imputation, and sensitivity analyses. Primary analyses and sensitivity analyses are compared to assess robustness of conclusions.

The statistical philosophy underlying this investigation emphasizes estimation rather than isolated hypothesis testing. The study seeks to quantify effect sizes, characterize uncertainty, develop prediction models, identify biological relationships, generate new hypotheses, and validate clinically meaningful imaging biomarkers. Accordingly, confidence intervals, calibration measures, discrimination metrics, decision-curve analyses, effect estimates, predictive performance measures, and clinical utility assessments receive greater emphasis than isolated P values. Multiplicity arising from numerous imaging biomarkers is addressed through predefined statistical methodologies appropriate to exploratory and confirmatory analyses.

Initial analyses summarize participant characteristics. Continuous variables are reported using mean, standard deviation, median, interquartile range, minimum, maximum, and distributional characteristics. Categorical variables are summarized using counts, percentages, frequency distributions, and cross-tabulations.

Visualization strategies include histograms, density plots, violin plots, box plots, heat maps, network diagrams, correlation matrices, trajectory plots, geographic mapping, body composition atlases, radiation dose maps, and three-dimensional cardiovascular visualizations.

Associations between imaging biomarkers and outcomes initially utilize conventional statistical methodology. Candidate models include linear regression, generalized linear models, logistic regression, ordinal regression, Poisson regression, negative binomial regression, mixed-effects models, generalized estimating equations, time-dependent Cox proportional hazards models, accelerated failure time models, competing-risk regression, flexible parametric survival models, restricted cubic splines, multistate models, and joint longitudinal-survival models. The specific methodology is selected according to the distribution and biological characteristics of each endpoint.

Traditional statistical methodology is complemented by machine learning analyses. Candidate approaches include random forests, gradient boosting, extreme gradient boosting, LightGBM, CatBoost, support vector machines, elastic net, neural networks, deep survival learning, graph neural networks, transformer architectures, multimodal fusion models, and ensemble learning. Performance is evaluated using repeated cross-validation, bootstrap validation, temporal validation, calibration analyses, discrimination metrics, and clinical decision analyses.

Every prediction model undergoes rigorous internal validation before interpretation. Methodologies include bootstrap resampling, nested cross-validation, Monte Carlo validation, temporal validation, repeated K-fold validation, leave-one-clinic-out validation, and leave-one-scanner-out validation. These approaches reduce optimism bias while providing realistic estimates of future model performance.

Whenever collaborative datasets become available, externally validated analyses are performed. External validation represents one of the highest priorities for future expansion because imaging biomarkers demonstrating reproducibility across institutions possess substantially greater translational potential. Collaborating institutions may contribute independent treatment planning CT examinations analyzed using identical methodologies. Performance consistency across institutions, scanners, patient populations, and treatment strategies is evaluated before future clinical implementation.

Oversight of the repository is provided by a multidisciplinary governance structure. Standing committees include:

• Scientific Steering Committee
• Data Governance Committee
• Imaging Core Committee
• Artificial Intelligence Committee
• Radiomics Working Group
• Radiation Dosimetry Committee
• Statistical Analysis Committee
• Clinical Outcomes Adjudication Committee
• Publications Committee
• External Scientific Advisory Board
• Trainee Education Committee
• Community Engagement Committee

Each committee possesses a written charter defining membership, responsibilities, meeting frequency, quorum, voting procedures, conflict-of-interest management, amendment procedures, and reporting structure. This framework ensures scientific rigor, regulatory compliance, methodological consistency, responsible stewardship of institutional data, and sustainable growth of the cardiovascular oncology imaging research enterprise.

The investigation is conducted in accordance with the ethical principles articulated in the Belmont Report, the Declaration of Helsinki, the ICH Guideline for Good Clinical Practice (ICH E6), the Common Rule (45 CFR 46), applicable provisions of the Health Insurance Portability and Accountability Act (HIPAA), relevant state and federal regulations, institutional policies, and all applicable requirements established by the Institutional Review Board.

Although this retrospective investigation involves minimal physical risk, the protocol recognizes that informational risks deserve equally rigorous consideration. Consequently, protections governing privacy, confidentiality, cybersecurity, artificial intelligence governance, and responsible data stewardship receive the same level of attention traditionally devoted to physical safety monitoring in interventional investigations.

The overall risk associated with participation in this retrospective investigation is anticipated to be minimal. Participants will not undergo additional clinical visits, laboratory testing, imaging examinations, radiation exposure, medication administration, biopsies, procedures, surveys, interviews, or behavioral interventions specifically for research purposes. No aspect of clinical management is modified based upon research participation.

The principal foreseeable risks relate to potential breaches of confidentiality, inadvertent disclosure of protected health information, inappropriate secondary use of clinical information, cybersecurity threats, or improper interpretation of observational findings. These risks are considered low because extensive safeguards described throughout this protocol substantially reduce their likelihood while limiting potential consequences should an unexpected event occur.

Protection of participant confidentiality represents one of the highest operational priorities of the study. Research identifiers independent of medical record numbers are assigned to every participant immediately following eligibility confirmation. Analytical datasets used for statistical analysis, artificial intelligence development, radiomics investigation, manuscript preparation, and educational activities preferentially utilize research identifiers rather than direct clinical identifiers whenever scientifically feasible.

Direct identifiers remain within secure institutional environments and are only accessible to authorized study personnel whose responsibilities specifically require access. Whenever possible, linkage between research identifiers and institutional identifiers is maintained within separate secure environments protected by role-based access controls. The minimum necessary identifiable information required to accomplish each approved research objective is utilized.

Because this study relies primarily upon existing clinical information, participant privacy is protected through strict adherence to institutional policies governing access to electronic health records and imaging archives. Only approved study personnel who have completed required institutional research training, privacy training, information security education, human subjects protection certification, and project-specific onboarding are granted access to study data. Access privileges are reviewed periodically throughout the duration of the investigation to ensure continued appropriateness. Investigators leaving the study have research access terminated promptly according to institutional policy.

The study complies fully with all applicable HIPAA regulations governing access, use, disclosure, transmission, storage, and retention of protected health information. Whenever possible, analytical datasets utilize de-identified information, coded information, or limited datasets according to approved regulatory pathways.

If identifiable protected health information is required during cohort identification, image linkage, longitudinal follow-up, or outcome ascertainment, such access occurs only within approved secure institutional environments under Institutional Review Board authorization.

The study anticipates requesting a Waiver of HIPAA Authorization for retrospective data collection because obtaining individual authorization from all eligible participants would be impracticable, would introduce substantial selection bias, and would significantly compromise scientific validity while providing minimal additional privacy protection beyond safeguards already incorporated into the protocol.

The study anticipates requesting a waiver of informed consent under applicable federal regulations governing minimal-risk retrospective research. Several factors support this request: the investigation involves no direct participant contact; no clinical intervention is performed; all clinical information was generated during routine medical care independent of the present investigation; requiring prospective consent from every historical participant would be impracticable given the size of the anticipated cohort, extended follow-up period, changes in residence, mortality, loss to follow-up, and elapsed time since treatment; restricting participation only to individuals successfully contacted would introduce substantial selection bias that could meaningfully compromise scientific validity, particularly for long-term cardiovascular outcomes; and extensive confidentiality protections substantially minimize informational risks associated with retrospective data use.

Because modern biomedical research increasingly depends upon digital infrastructure, cybersecurity represents an essential component of participant protection. Research data reside only within institutionally approved computing environments employing encryption during transmission and storage, multifactor authentication where available, role-based authorization, audit logging, routine security monitoring, automated backup procedures, disaster recovery planning, vulnerability management, endpoint protection, and institutional cybersecurity oversight.

Portable storage devices are not used for long-term research data storage except under specifically approved encrypted institutional procedures. Cloud computing resources may be utilized only if approved by institutional information security offices and compliant with applicable regulatory requirements. Artificial intelligence development environments operate under identical security standards as traditional analytical environments.

Artificial intelligence development within this protocol is intended exclusively for scientific investigation and future clinical decision support. Algorithms generated during this study do not independently make clinical decisions regarding patient diagnosis, treatment planning, medication management, prognosis, or radiation delivery. Instead, AI outputs function as investigational research tools subject to expert clinical interpretation.

Algorithm development emphasizes transparency, explainability, reproducibility, calibration, fairness, uncertainty estimation, external validation, and continuous quality assessment before any future consideration of clinical implementation. Models demonstrating unacceptable bias, instability, poor calibration, inadequate generalizability, or inconsistent performance across clinically relevant populations undergo additional refinement before further investigation.

Because imaging analyses performed within this retrospective investigation are conducted for research rather than clinical purposes, individual research findings generally will not be returned to participants. However, should future analyses identify findings that are analytically valid, clinically actionable, medically important, verified using clinically approved methodologies, and accompanied by an approved institutional return-of-results process, future protocol amendments may establish mechanisms for appropriate clinical confirmation and participant notification. At present, no investigational imaging measurements generated solely for research purposes will directly influence clinical management.

Individual participants should not expect direct clinical benefit from participation in this retrospective investigation. The principal anticipated benefits are societal and scientific. These include improved understanding of cardiovascular toxicity associated with cancer therapy, enhanced cardiovascular risk prediction before radiation treatment, development of automated imaging biomarkers, advancement of artificial intelligence methodologies, optimization of radiation treatment planning, improved survivorship strategies, more personalized cardiovascular monitoring, and generation of preliminary data supporting future prospective clinical investigations.

The imaging repository established through this protocol may ultimately contribute to clinical decision-support systems capable of improving cardiovascular outcomes for future patients with cancer.

The investigative team supports responsible scientific collaboration while recognizing the ethical obligation to protect participant privacy. Future sharing of de-identified analytical datasets with collaborating investigators may occur following appropriate Institutional Review Board approval, executed data use agreements, institutional authorization, scientific review, and governance committee approval. Priority is given to collaborative investigations that advance cardiovascular oncology, improve reproducibility, validate imaging biomarkers, enhance artificial intelligence methodology, or facilitate multicenter implementation. All external collaborations comply with applicable institutional, state, federal, and international regulatory requirements.

Scientific findings generated through this investigation are disseminated through peer-reviewed publications, scientific meetings, educational programs, collaborative workshops, trainee mentorship activities, institutional seminars, national conferences, international cardiovascular oncology meetings, and future implementation science initiatives.

Publication decisions prioritize scientific rigor, methodological transparency, reproducibility, appropriate acknowledgment of collaborators, adherence to authorship standards, and timely dissemination of findings. Negative findings, neutral findings, validation studies, methodological investigations, and quality improvement discoveries receive equal scientific consideration because each contributes meaningfully to advancement of cardiovascular oncology.

This retrospective investigation represents the foundational phase of a much larger scientific program rather than an isolated observational study. The repository established through this protocol is intentionally designed to evolve continuously through incorporation of additional imaging examinations, longer follow-up, prospective enrollment, biospecimen collection, molecular phenotyping, wearable technologies, patient-reported outcomes, implementation science, pragmatic clinical trials, multicenter collaboration, and learning health system methodologies.

Over time, the integrated cardiovascular oncology platform developed through the collaboration among Dartmouth Radiation Oncology, Heart Innovation and Equity Research (HIER) Institute, and Heart Spark Research & Innovation Institute is intended to become an internationally recognized resource for precision cardiovascular oncology research.

Ultimately, the scientific aspiration extends beyond publication alone. The long-term goal is to enable every radiation treatment planning CT examination to contribute simultaneously to cancer care, cardiovascular prevention, artificial intelligence development, scientific discovery, education, and continuous improvement of survivorship care.

Manuscripts arising from this platform will be prospectively planned through a joint publication committee representing all partner institutions. Authorship will follow ICMJE criteria with explicit contribution statements. Trainee first-authorships will be actively supported. De-identified datasets, statistical code, and algorithm artifacts will be shared through appropriate repositories in accordance with NIH data-sharing policy, journal requirements, and executed agreements. Pre-registration of sub-studies will be encouraged where methodologically appropriate.

The following appendices accompany the master protocol and will be maintained as living documents by the Data Coordinating Center:

• A. Case Report Forms
• B. REDCap Data Dictionary
• C. Variable Dictionary — Clinical
• D. Variable Dictionary — Imaging
• E. Variable Dictionary — Dosimetry
• F. Cardiovascular Outcome Definitions
• G. Oncologic Outcome Definitions
• H. Adverse Event Definitions
• I. Image Processing SOP
• J. Cardiac Segmentation SOP
• K. Radiation Dosimetry Manual
• L. AI/ML Analysis Plan
• M. Radiomics Analysis Plan
• N. Statistical Analysis Plan
• O. Data Management Plan
• P. Data Security Plan
• Q. HIPAA Waiver Justification
• R. Waiver of Consent Justification
• S. Governance Plan
• T. Investigator Responsibilities
• U. Publication Plan
• V. Data Sharing Policy
• W. Regulatory Binder
• X. References

A companion Prospective Master Protocol, AI/ML Protocol, Radiomics Protocol, Imaging Core Manual, Dosimetry Manual, Data Coordinating Center Manual, REDCap Data Dictionary, Case Report Forms, Standard Operating Procedures, Regulatory Binder, HIPAA and Waiver Documentation, Investigator Manual, Publication Manual, Data Sharing Manual, and Governance Manual are maintained under the same umbrella framework and are available on request.