HIER Institute × Dartmouth

Robins Lab Collaboration

A proposed multi-institutional collaboration with Dr. Marthony L. Robins, PhD, Assistant Professor of Radiology and Medical Physics at the Geisel School of Medicine at Dartmouth, integrating quantitative CT physics, radiomics, and cardiovascular oncology imaging science.

Proposal notice: Everything on this page is being proposed by the HIER Institute for consideration by the Robins Lab and collaborating partners. Aims, endpoints, timelines, governance, and funding pathways are draft proposals — not yet finalized — and remain subject to review.

Robins Lab website ↗Umbrella IRB pathwayProuty Pilot · Hitchcock Foundation

Featured proposal · exploratory collaboration

A best-in-class combined cardiovascular predictor for patients on radiation oncology.

Using cancer cases acquired from 2010–2025 under Dr. Robins's umbrella IRB, including radiation planning CT images, associated metadata, pathology, and EHR access as needed, we will build and compare cardiovascular disease predictors specifically for cancer survivors undergoing radiation oncology and benchmark them against predictors that rely on echocardiography, clinical data, and demographics alone.

Model A — CT only. Prediction from radiation planning CT (CAC, EAT, aortic calcium, myocardial texture, cardiac substructures).
Model B — ECG only. Prediction from 12-lead ECG images / waveforms acquired around the RT course.
Model C — CT + ECG fusion. Multimodal transformer combining planning CT and ECG.
Model D — full multimodal. CT + ECG + echocardiography + other available cardiovascular imaging + clinical data and demographics.
Comparator — non-CT baseline. Best predictor achievable from echocardiography, clinical data, and demographics alone, without any planning-CT features.

Cohort

Cancer patients treated from 2010–2025 with a radiation planning CT and downstream cardiovascular follow-up.

Data access

CT images and metadata under the existing umbrella; EHR structured and unstructured data (labs, notes, cardiovascular imaging) as needed; enterprise informatics resources under evaluation.

Support

Two experienced biostatisticians engaged through the Robins lab, medical student engagement, and potential application for seed funding via institutional pilot mechanisms.

Endpoints (each analyzed individually and as a composite)

Coronary artery disease (CAD) — incident CAD by diagnosis codes, imaging, or revascularization.
Myocardial infarction (MI) — Fourth Universal Definition, type 1/2 distinguished.
Stroke — ischemic and hemorrhagic, imaging-confirmed.
Heart failure — incident HF diagnosis and HF hospitalization (first and recurrent).
Cardiomyopathy — incident cardiomyopathy (ischemic, non-ischemic, cancer therapy–related), including asymptomatic LVEF decline and GLS-based dysfunction.
Atrial fibrillation — incident AF by ECG, monitor, or discharge diagnosis.
Cardiovascular death — adjudicated cause-specific mortality.
Coronary revascularization — PCI or CABG.
Composite MACE — any of the above, with cardiac-substructure dose–response tracked as a secondary endpoint.

Primary comparison: discrimination and calibration of Model D versus the non-CT comparator, with pre-specified sub-analyses by cancer site, mean heart dose, and age. Next step: Zoom to align expectations and confirm umbrella IRB scope.

Analytic framework (adapted from prior reviewer feedback)

At-risk denominators. For every individual endpoint, patients with that condition prior to cancer diagnosis are excluded from the incident-outcome analysis of that same condition; denominators are reported per endpoint rather than as a single full-cohort n, and recurrence/exacerbation of pre-existing disease is analyzed separately.
Time-to-event, not cumulative counts. Kaplan–Meier and cause-specific Cox proportional hazards models are the primary framework, with cardiovascular death treated as a competing risk (Fine–Gray subdistribution hazards) for non-fatal endpoints. Follow-up windows are standardized from index CT date.
Multivariable adjustment. Given expected baseline imbalance between imaged subgroups, all primary comparisons are adjusted at minimum for age, sex, prior cardiovascular disease burden, comorbidity index, cancer site and stage, treatment class (anthracycline, HER2, immune checkpoint, radiation), mean heart dose, and cardioprotective medications. Unadjusted comparisons are reported only as descriptive context.
Oncologic confounders. Cancer type, stage, tumor burden, and cancer-specific mortality are included as covariates and as effect modifiers in pre-specified interaction analyses.
Equity and access. Screening-rate and outcome disparities across race/ethnicity, rurality, and area-level socioeconomic status are examined jointly — separating the marker-of-access signal from any independent biological signal.
Multiplicity and power. Benjamini–Hochberg control across endpoint families, with pre-specified post-hoc power reporting for subgroups; negative findings in small subgroups are reported as exploratory only.
Sensitivity analyses. Landmark analyses at 6, 12, and 24 months; short-, medium-, and long-term follow-up strata; propensity-score adjustment and E-values for unmeasured confounding; and a healthcare-engagement sensitivity model (routine cardiology follow-up, ECG availability) to test whether imaging findings carry signal beyond being a marker of higher-acuity care.
Generalizability. Single-system results are framed accordingly; external validation is prespecified via the Model C/D transfer to a second site under the umbrella IRB.

Companion study proposal

Breast MRI phenotype → treatment exposure → ECG change → cardiovascular outcome

A parallel, immediately feasible design that requires no new cardiac MRI AI model: the existing breast MRI supplies a one-time baseline phenotype, same-patient ECGs supply the cardiac electrical phenotype, cancer therapies supply the exposures, and echocardiographic or adjudicated clinical events supply the endpoints. The dedicated page details feasibility tiers by available MRI material, seven candidate ECG studies, the recommended first study, and a feasibility inventory to run before the hypothesis is locked.

Read the ECG phenotyping study proposal →

Context from Dr. Robins

Dr. Robins has confirmed access to two experienced biostatisticians with extensive clinical and translational experience, an existing umbrella IRB protocol that likely covers much of the envisioned data range (making an amendment more feasible than a de novo submission), and two internal pilot funding mechanisms — including the Prouty Pilot Grant Program and Hitchcock Foundation grants — suitable for seed funding of feasibility studies and preliminary data generation.

Overall objective

To develop and validate an interpretable, clinically actionable cardiovascular risk model for radiation-treated cancer survivors using routinely acquired radiation planning CT, ECG, and clinical data.

Central hypothesis

Quantitative cardiovascular phenotypes extracted from radiation planning CT provide independent prognostic information beyond standard clinical variables, and their integration with ECG data will improve identification of cancer survivors at elevated cardiovascular risk.

Specific Aims

Define the prognostic cardiovascular phenotype contained within radiation planning CT.

HypothesisA reproducible and parsimonious set of planning CT features will independently predict subsequent cardiovascular events and reveal clinically meaningful dose–response relationships.

Approach

Extract coronary artery calcium (Agatston and volumetric), epicardial adipose tissue volume and attenuation, thoracic aortic calcium, myocardial texture, and cardiac substructure characteristics from planning CTs.
Compute per-substructure radiation dose metrics (mean, near-max, and volume thresholds) with EQD2 correction.
Adjust for baseline cardiovascular risk (ASCVD, SCORE2), cancer characteristics (site, stage, histology), and treatment exposures (systemic therapy, mean heart dose).
Determine which features independently predict the prespecified major cardiovascular event composite and each individual component using cause-specific Cox and Fine–Gray competing-risks models.

Deliverables

Parsimonious planning CT phenotype with feature-level effect estimates and reproducibility metrics.
Prespecified dose–response curves for cardiac substructures with uncertainty bands.
Manuscript targeting JACC: CardioOncology or Int J Radiat Oncol Biol Phys.

Develop and temporally validate an interpretable multimodal cardiovascular risk model.

HypothesisA parsimonious model integrating CT, ECG, and clinical variables will provide greater predictive accuracy and clinical utility than models based on clinical information alone.

Approach

Prespecified head-to-head comparison of CT-only, ECG-only, CT + ECG, and CT + ECG + clinical models against a clinical comparator (ASCVD / SCORE2 plus oncologic exposures).
Select the smallest model that preserves strong performance; report feature attributions and case-level explanations.
Evaluate discrimination (time-dependent AUC, C-index), calibration (intercept, slope, calibration plots), reclassification (NRI, IDI), and decision-curve net benefit.
Temporal held-out validation using a prespecified index-date split within the Dartmouth Health 2010–2025 cohort.

Deliverables

Locked multimodal model with published TRIPOD+AI report and model card.
Web calculator and nomogram integrated into the HIER dashboard.

★

Future directions

Establish the transportability, equity, and clinical actionability of the final locked model.

HypothesisThe model will retain useful performance across major clinical and demographic populations while identifying specific settings that require recalibration or additional data.

Directions

Evaluate calibration, discrimination, and net benefit across cancer sites, radiation dose strata, age, sex, race and ethnicity, rurality, and socioeconomic strata.
Validate the locked model in an independent second-site cohort once data access and regulatory approvals are formally secured.
Define clinically meaningful risk groups and candidate surveillance thresholds for prospective evaluation.
Assess clinical utility and robustness of the final model across prespecified cancer, treatment, and demographic subgroups.

Bottom-line structureBiological and imaging discovery → multimodal model development → validation and clinical translation. Time-to-event framing, competing-risks methods, covariate adjustment, multiplicity control, sensitivity analyses, missing-data procedures, and power analyses sit within the Approach and support all aims and future directions.

Endpoint Definitions

Each endpoint is defined by clinical criteria, an outcome time window relative to the index planning CT, primary data sources, and an adjudication rule. The endpoint set is aligned with the cardio-oncology outcomes and cohort infrastructure described in AHA 2024 Scientific Statement (10.1161/CIR.0000000000001417), American Heart Journal Plus 2022 (10.1016/j.ahjo.2022.100094), and JAHA 2024 (10.1161/JAHA.123.030363), while keeping formal endpoint definitions anchored to sources that actually publish standardized endpoint rules. See the Referenced Cardiovascular Outcomes panel below for the exact endpoint list drawn from each source.

Coronary artery disease (CAD)

event_type=cad · icd10 (I25.x) · coronary_revasc · ct_cac_agatston

Time window: Index CT → end of follow-up; prior CAD excluded from the incident-CAD at-risk denominator.
Data sources: EHR problem list and encounter diagnoses, cardiology consult, cath report, coronary CTA, planning-CT CAC.
Adjudication rule: Incident CAD defined by first documented ICD-10 I25.x, obstructive disease on angiography or CTA, or first coronary revascularization. Analyzed individually and as a component of composite MACE.

Myocardial infarction (MI)

mi · event_type=mi · hs_troponin_ng_l

Time window: Any incident MI from 30 days pre-CT through end of follow-up.
Data sources: EHR (troponin trajectory, ECG, cath), discharge summary, cardiology consult note.
Adjudication rule: Fourth Universal Definition of MI; type 1 vs type 2 distinguished. First event contributes to time-to-first analyses; recurrent MI captured for count-based sensitivity models.

Stroke

stroke · event_type=stroke

Time window: Index CT → end of follow-up.
Data sources: Neurology consult, brain MRI/CT report, discharge summary.
Adjudication rule: Ischemic and hemorrhagic strokes counted; TIA excluded from the primary endpoint but tracked as a secondary outcome. Imaging confirmation required.

Heart failure

hf_hospitalization · event_type=hf · icd10 (I50.x) · nt_probnp_pg_ml · bnp_pg_ml · echo_lvef_pct

Time window: Index CT → end of follow-up; first and recurrent HF hospitalizations.
Data sources: Inpatient discharge summary with HF as primary or contributing diagnosis, natriuretic peptide trajectory, echo LVEF/GLS, outpatient HF diagnosis codes.
Adjudication rule: Incident HF captured from first HF diagnosis or first HF hospitalization (acute decompensation requiring intravenous diuresis or inotropes with a discharge diagnosis of HF). HFrEF vs HFpEF subclassified when LVEF is available within 90 days.

Atrial fibrillation

new_afib · ecg_rhythm · icd10 (I48.x)

Time window: Index CT → end of follow-up; prior AF excluded from the incident-AF at-risk denominator.
Data sources: 12-lead ECG, ambulatory monitor, discharge summary, problem list.
Adjudication rule: First documented AF or atrial flutter ≥30 seconds. Analyzed as an individual endpoint and as a prespecified secondary component alongside composite MACE.

Cardiovascular death

cv_death · event_type=cv_death · overall_survival_status

Time window: Index CT → end of follow-up; landmarks at 1, 3, and 5 years.
Data sources: Death certificate, EHR discharge/mortality note, SSDI linkage, tumor registry.
Adjudication rule: Cause-specific mortality adjudicated by two reviewers; disagreements resolved by tiebreaker (see 09_adjudication.csv). Non-cardiovascular death is a competing event for all non-fatal endpoints.

Coronary revascularization

coronary_revasc · event_type=coronary_revasc

Time window: Index CT → end of follow-up.
Data sources: Cath lab report, operative note, procedure code (CPT/ICD-PCS).
Adjudication rule: Percutaneous coronary intervention or coronary artery bypass grafting; staged procedures within 30 days count as a single event for time-to-first analyses.

Composite MACE (primary)

mace_composite (derived from cv_death, mi, stroke, hf, coronary_revasc; CAD and AF tracked in parallel)

Time window: Index CT → time to first component event.
Data sources: Union of the individual endpoint sources above.
Adjudication rule: Time-to-first-event; recurrent events analyzed in a prespecified secondary count model. Non-cardiovascular death is a competing event.

Cardiomyopathy

cardiomyopathy_incident · cardiomyopathy_subtype · echo_lvef_pct · echo_gls_pct

Time window: Index CT → end of follow-up; imaging assessed at 3, 6, 12, 24, and 60 months.
Data sources: Echocardiography (biplane Simpson LVEF, GLS), cardiac MRI when available, discharge diagnoses (ICD-10 I42.x).
Adjudication rule: Incident cardiomyopathy classified as ischemic, non-ischemic, or cancer therapy–related. Includes symptomatic HF and asymptomatic LVEF decline ≥10 percentage points to <53%, or ≥15% relative decline in GLS. Analyzed individually and as a component of composite MACE.

Cancer therapy–related cardiac dysfunction (CTRCD)

echo_lvef_pct · echo_gls_pct · valve_dysfunction · pericardial_disease

Time window: Index CT → 5 years; per-visit at 3, 6, 12, 24, and 60 months.
Data sources: Echocardiography (biplane Simpson LVEF, GLS), cardiac MRI when available.
Adjudication rule: Aligned with AHA and cardio-oncology imaging definitions: symptomatic HF, or asymptomatic LVEF decline ≥10 percentage points to <53%, or ≥15% relative decline in GLS. Pericardial and valvular events tracked as separate secondary endpoints.

Cardiac-substructure dose–response (secondary)

lad_mean_gy · left_main_mean_gy · lv_mean_gy · la_mean_gy · san_mean_gy · avn_mean_gy · mean_heart_dose_gy

Time window: Index CT → time to any cardiovascular endpoint above.
Data sources: TPS DVH extraction on auto-segmented substructures on the planning CT.
Adjudication rule: Restricted cubic splines with clinically prespecified knots; report thresholds with bootstrap uncertainty and benchmark against QUANTEC/RTOG.

Referenced Cardiovascular Outcomes and Source Roles

The individual cardiovascular outcomes drawn from each cited source. HIER's endpoint set (above) is a superset that maps one-to-one back to these references.

Reference 1

AHA 2024 Scientific Statement — Cardiovascular Outcomes in Cancer Patients: Standardized Definitions (Circulation, 10.1161/CIR.0000000000001417)

Cardiovascular death
Major adverse cardiovascular events (MACE): CV death, nonfatal myocardial infarction, nonfatal stroke
Heart failure hospitalization or urgent heart-failure visit
Cancer therapy–related cardiac dysfunction (CTRCD)
Arrhythmias: atrial fibrillation, ventricular tachycardia, ventricular fibrillation, heart block/bradyarrhythmia
Coronary revascularization (PCI or CABG)
Acute limb ischemia and other arterial thromboembolism
Venous thromboembolism (DVT and pulmonary embolism)
Bleeding events, including hemorrhagic stroke
Pericardial disease (effusion, tamponade, constrictive pericarditis)
Valvular heart disease
Hypertension and hypotension events

Reference 2

American Heart Journal Plus 2022 — Establishing an interdisciplinary research team for cardio-oncology AI informatics precision and health equity (10.1016/j.ahjo.2022.100094)

Accurate source role: team science, AI informatics, health equity, and a cancer-survivor epidemiology cohort. It should not be treated as a formal endpoint-definition paper.

All-cause mortality
Cardiovascular mortality
Incident heart failure
Incident atrial fibrillation
Ischemic heart disease events, including myocardial infarction
Cerebrovascular events (stroke and TIA)
Peripheral arterial disease events
Venous thromboembolism
Cardiovascular hospitalization (composite)
Incident hypertension, diabetes, and dyslipidemia as cardiovascular comorbidities

Reference 3

JAHA 2024 — Machine-Learning Prediction of Cardiovascular Disease in Cancer Survivors (10.1161/JAHA.123.030363)

Composite cardiovascular disease (CVD): the primary predicted outcome, comprising incident coronary artery disease, myocardial infarction, and stroke
Coronary artery disease (CAD)
Myocardial infarction (MI)
Stroke
Heart failure
Atrial fibrillation

Outcomes are mapped to each source's role; the American Heart Journal Plus source is retained for cohort infrastructure and the user-provided outcome extract, not for formal endpoint adjudication rules. HIER's Endpoint Definitions section above encodes these into the data dictionary (see event_type enum in 04_outcomes_longitudinal.csv).

Analysis Modules

Each module reports both individual endpoints and prespecified composites, uses consistent per-endpoint at-risk denominators, and handles competing risks explicitly.

Module A — Individual endpoints

•Kaplan–Meier estimates and cause-specific Cox proportional hazards for CV death, MI, stroke, HF hospitalization, and coronary revascularization.
•Per-endpoint at-risk denominators: patients with the same condition before the index CT are excluded from the incident analysis of that endpoint.
•Non-cardiovascular death treated as a competing event via Fine–Gray subdistribution hazards.
•Report HR, 95% CI, and c-index per model.

Module B — Composite MACE

•Primary composite: time to first CV death, MI, stroke, HF hospitalization, or coronary revascularization.
•Cause-specific and Fine–Gray models fit in parallel; concordance in direction is a prespecified robustness check.
•Recurrent-event analysis (Andersen–Gill) as a secondary count model.
•Landmark analyses at 6, 12, and 24 months to reduce immortal-time bias.

Module C — Secondary AF and CTRCD

•Incident AF analyzed on the AF-naive at-risk denominator.
•CTRCD modeled as a time-to-event outcome using serial echo LVEF and GLS, aligned with published cardio-oncology definitions.
•Pericardial disease and valvular dysfunction reported separately with explicit denominators.

Module D — Cardiac-substructure dose–response

•Per-substructure mean and near-max dose entered as restricted cubic splines.
•Interaction terms between substructure dose and baseline CT phenotype (CAC, EAT, aortic calcium).
•Report absolute risk differences at prespecified dose thresholds and bootstrap uncertainty bands.
•Benchmark against QUANTEC and RTOG whole-heart constraints.

Sensitivity Analyses

Prespecified robustness checks for the Robins Lab framework. Each row lists the linked 00_data_dictionary.csv fields used to execute the analysis.

Sensitivity theme	Specification	Dictionary fields
Alternative model specifications	Cause-specific Cox vs Fine–Gray; parametric AFT (Weibull); penalized Cox (elastic net); gradient-boosted survival for nonlinear checks.	event_type · event_date · overall_survival_status · last_followup_date
Missingness handling	Multiple imputation by chained equations (m=20) for covariates; complete-case analysis as a comparator; missing-indicator method for structurally missing fields.	ldl_c_mg_dl · hdl_c_mg_dl · hba1c_pct · echo_lvef_pct · echo_gls_pct
Unmeasured confounding	E-values on primary estimates; propensity-score adjustment and matching on ASCVD, SCORE2, and cancer stage.	ascvd_10yr_pct · score2 · framingham_pct · ajcc_stage · concurrent_systemic
Healthcare engagement	Adjust for routine cardiology follow-up, ECG availability, and echo availability to isolate imaging signal from marker-of-care effects.	echo_available · ccta_available · cardiac_mri_available · ecg_rhythm
Landmark and time windows	Landmark analyses at 6, 12, and 24 months; short (<2 y), medium (2–5 y), and long (>5 y) follow-up strata.	scan_date · rt_start_date · event_date · last_followup_date
Dose reclassification	Recompute all dose–response models with EQD2 correction (α/β = 3 for heart) and DIBH vs free-breathing strata.	eqd2_gy · mean_heart_dose_gy · breath_hold · lad_mean_gy · lv_mean_gy
Harmonization robustness	Repeat all primary models on pre- and post-harmonization biomarker values; report ICC shift and effect-estimate stability.	cac_agatston · eat_volume_ml · eat_mean_hu · thoracic_aortic_calcium_score · recon_kernel · iterative_recon_level
Subgroup transportability	Refit and recalibrate in prespecified subgroups: cancer site, sex, age band, race/ethnicity, rurality, and mean heart dose tertile.	cancer_diagnosis · sex · age_at_rt_start · race · ethnicity · mean_heart_dose_gy
Multiplicity control	Benjamini–Hochberg FDR across endpoint families; family-wise error control (Holm) for the primary composite.	mace_composite · cv_death · mi · stroke · hf_hospitalization · coronary_revasc

Comprehensive Data Points

The following elements support all three aims and would be scoped against Dr. Robins's umbrella IRB during the amendment feasibility review.

Imaging (Planning CT + adjunct)

•Non-contrast radiation therapy planning CT (4DCT average, free-breathing, and breath-hold when available)
•Contrast-enhanced planning CT when acquired; diagnostic chest CT when available
•DICOM headers: manufacturer, model, kernel, slice thickness, kVp, mAs/CTDIvol, pitch, iterative reconstruction level
•RT structure sets: heart, cardiac substructures, great vessels, lungs, esophagus, targets
•RT dose grids and plan files for DVH extraction and substructure re-analysis
•Cardiac substructure auto-segmentations (LAD, LM, RCA, LCx, LV, RV, LA, RA, SAN, AVN, pericardium, EAT)
•Quantitative biomarkers: Agatston CAC, volume CAC, EAT volume + mean HU, thoracic aortic calcification score, myocardial mean HU, myocardial radiomic features
•Available cardiac MRI, echocardiography, coronary CTA, and nuclear perfusion where clinically obtained

Clinical & demographic

•Age, sex, self-reported race/ethnicity, height, weight, BMI, BSA
•Cancer diagnosis, laterality, stage (AJCC), histology, molecular markers
•Cardiovascular risk factors: hypertension, diabetes, dyslipidemia, tobacco use (pack-years), family history of premature CAD
•Baseline ASCVD 10-year risk, SCORE2/SCORE2-OP, Framingham score
•Pre-existing cardiovascular disease (CAD, HF, arrhythmia, valvular disease, prior revascularization)
•Menopausal status; pregnancy history where relevant

Oncologic treatment exposures

•Radiation modality (3D-CRT, IMRT/VMAT, proton, SBRT), fractionation, total dose, EQD2
•Whole-heart Dmean/Dmax and substructure DVH metrics (Dmean, D0.03cc, V5, V20, V25, V30, V40)
•Concurrent and sequential systemic therapy: anthracyclines (cumulative doxorubicin-equivalent dose), HER2-targeted agents, VEGF/TKIs, immune checkpoint inhibitors, endocrine therapy, 5-FU/capecitabine
•Surgical history relevant to cardiac exposure

Longitudinal outcomes

•MACE (composite): cardiovascular death, MI, stroke, hospitalization for heart failure, coronary revascularization
•Individual endpoints with adjudicated dates
•Incident arrhythmia (atrial fibrillation, ventricular arrhythmia), pericardial disease, valvular dysfunction
•Cancer-specific survival, overall survival, recurrence-free survival
•Cardio-oncology clinic encounters, cardioprotective medications initiated

Biomarkers & laboratory

•Lipid panel (LDL-C, HDL-C, triglycerides, non-HDL-C, Lp(a) when available)
•HbA1c, fasting glucose
•hs-CRP, NT-proBNP or BNP, high-sensitivity troponin (baseline and surveillance)
•eGFR, creatinine, liver function, complete blood count
•Optional biobanking for future proteomic/metabolomic panels

Functional and patient-reported

•Baseline ECG (rhythm, intervals, LVH criteria)
•Baseline echocardiography (LVEF, GLS, diastolic function, chamber volumes, valve function)
•6-minute walk / functional capacity when available
•Patient-reported outcomes: PROMIS Physical Function, dyspnea, fatigue, quality of life

Physics / QA metadata (Robins Lab lead)

•Scanner calibration logs, HU stability records
•Reconstruction algorithm version, iterative reconstruction strength
•Dose reduction settings, tube current modulation profile
•Phantom scans (anthropomorphic thorax with calcium and adipose inserts) matched to clinical acquisitions
•Vendor firmware / software version at time of acquisition

IRB pathway

Evaluate Dartmouth umbrella IRB scope against Aim 1–3 data requirements.
File a targeted amendment covering multi-institutional data sharing, harmonization, and outcome linkage.
Execute a Data Use Agreement between HIER Institute and Dartmouth; NIH-compliant Data Management and Sharing Plan.
Coordinate with HIER IRB reliance / single-IRB (sIRB) mechanism where eligible.

Pilot funding strategy

Prouty Pilot Grant — feasibility of Aim 1 harmonization and Aim 2 phenotype derivation.
Hitchcock Foundation — seed funding for Aim 3 substructure dose–response analysis.
Preliminary data package for subsequent NIH R01, NCI U01, or DoD BCRP submission.
Statistician effort covered under pilot budget lines (two Dartmouth biostatisticians identified).

IRB & Amendment Status Tracker

Live status of umbrella IRB coverage, required amendments, and review and funding milestones for the HIER × Robins Lab collaboration. Updated by the study coordinator; ambient status only, not a regulatory record.

CompletePendingBlocked

Umbrella IRB coverage

Retrospective planning CT DICOM (Dartmouth)

Likely covered

Dr. Robins reports existing umbrella covers most imaging data.

RT dose grids and structure sets

Likely covered

Included under existing radiation oncology umbrella.

Multi-institutional data sharing (HIER ↔ Dartmouth)

Amendment needed

Amendment required for outbound sharing and DUA execution.

Cardiovascular outcome linkage (EHR + adjudicated MACE)

Amendment needed

Requires explicit outcome ascertainment language.

Retrospective biospecimen linkage

TBD

Feasibility review by Dartmouth IRB pending.

Prospective enrollment / re-consent

Not covered

Out of scope for pilot; separate protocol.

Required amendments

AMD-01

Multi-institutional data sharing + DUA

Owner: Robins Lab / HIER Reg. Affairs · Target: Q1

Pending

AMD-02

Outcome ascertainment (adjudicated MACE)

Owner: HIER Endpoints Committee · Target: Q1

Pending

AMD-03

Cardiac substructure auto-segmentation + secondary use

Owner: Robins Lab · Target: Q2

Pending

AMD-04

Phantom + in-vivo harmonization dataset release

Owner: Robins Lab CT Physics · Target: Q2

Pending

AMD-05

sIRB reliance agreement (HIER ↔ Dartmouth CPHS)

Owner: HIER Reg. Affairs · Target: Q1–Q2

Pending

Review & funding milestones

MS-01

Introductory call — scope, co-PI, statistician alignment

Owner: PIs · Target: Week 1

Pending

MS-02

Circulate three-aim brief to Dartmouth biostatisticians

Owner: HIER · Target: Week 2

Pending

MS-03

Umbrella IRB scope confirmation

Owner: Robins Lab · Target: Week 3

Pending

MS-04

Prouty Pilot Grant submission

Owner: Co-PIs · Target: Next cycle

Pending

MS-05

Hitchcock Foundation LOI

Owner: Co-PIs · Target: Q2

Pending

MS-06

Shared harmonized planning-CT sandbox live

Owner: HIER Informatics · Target: Q2

Pending

MS-07

Preliminary data package for NIH R01 / DoD BCRP

Owner: Co-PIs · Target: Q4

Pending

Study-specific retrospective protocol

Read the full retrospective IRB protocol for this study

A companion protocol tailored to the endpoints, cohort, and data sources above — cohort definition, data sources, variables, endpoints, statistical analysis plan, privacy safeguards, waiver justification, umbrella IRB pathway, timeline, and study team.

Open study-specific IRB protocol →Master retrospective protocol

Data Dictionary & Submission Templates

One CSV per data domain. Each template ships with a header row plus one EXAMPLE_ row showing units and formats — delete that row before returning the file. UTF-8, comma-delimited, dates in ISO-8601. See the README for encoding and identifier conventions.

Master Data Dictionary

Variable-level dictionary for every field across templates 01–10: type, units, allowed values, PHI flag, derivation, source, and validation rule.

Download CSV Preview ↗

00_data_dictionary.csv

Imaging (Planning CT + adjunct)

Per-scan DICOM metadata, RT UIDs, and derived cardiovascular biomarkers (CAC, EAT, aortic calcium, myocardial HU).

Download CSV Preview ↗

01_imaging.csv

Clinical & Demographics

Per-subject demographics, anthropometrics, oncologic diagnosis, CV risk factors, and baseline risk scores.

Download CSV Preview ↗

02_clinical_demographics.csv

Oncologic Treatment Exposures

Per-plan RT dose and fractionation, whole-heart and substructure DVH metrics, and concurrent systemic therapy.

Download CSV Preview ↗

03_oncologic_treatment.csv

Longitudinal Outcomes

Adjudicated MACE, individual endpoints, arrhythmia, pericardial and valvular events, and survival status.

Download CSV Preview ↗

04_outcomes_longitudinal.csv

Biomarkers & Laboratory

Lipids, HbA1c, hs-CRP, NT-proBNP/BNP, hs-troponin, renal and hepatic panels at each timepoint.

Download CSV Preview ↗

05_biomarkers_labs.csv

Functional & Patient-Reported

ECG, echo (LVEF/GLS), 6MWT, and PROMIS-based patient-reported outcomes.

Download CSV Preview ↗

06_functional_pro.csv

Physics / QA Metadata

Scanner make/model, kernel, iterative reconstruction strength, dose reduction, phantom QA, and HU stability logs.

Download CSV Preview ↗

07_physics_qa.csv

Medications

Concomitant medications with RxNorm codes, drug class, indication, and cumulative anthracycline (doxorubicin-equivalent) exposure.

Download CSV Preview ↗

08_medications.csv

Adjudication Log

Endpoints-committee two-reviewer adjudication with tiebreaker, source documents reviewed, and turnaround metrics.

Download CSV Preview ↗

09_adjudication.csv

Biospecimen Inventory

Optional biobank inventory: specimen type, processing delay, freeze–thaw, aliquot count, and linked scan and plan IDs.

Download CSV Preview ↗

10_biospecimen.csv

Open README ↗Download master dictionary

Searchable Data Dictionary

Live view of 00_data_dictionary.csv. Filter by domain, type, required status, and free text to find any variable across the 11 templates.

Domain

Type

Required only

Loading…

Variable	Domain	Type	Units	Req.	PHI	Allowed / range

Role-Based Data Access Matrix

Least-privilege access per collaborator role, per data domain, per governance step. Access is granted only after the corresponding governance step is complete (see the flow at the bottom of this page).

Identifiers (crosswalk)

PI / co-PIAudit

CoordinatorEdit

CT physicist—

Biostatistician—

Analyst / ML—

Adjudicator—

Reg. affairsAudit

External viewer—

DUA executed + honest broker

Planning CT + DICOM

PI / co-PIView

CoordinatorEdit

CT physicistEdit

BiostatisticianView

Analyst / MLView

Adjudicator—

Reg. affairs—

External viewer—

IRB amendment + de-id complete

RT plans / structures / dose

PI / co-PIView

CoordinatorEdit

CT physicistView

BiostatisticianView

Analyst / MLView

Adjudicator—

Reg. affairs—

External viewer—

IRB amendment

Clinical & demographics

PI / co-PIView

CoordinatorEdit

CT physicist—

BiostatisticianExport

Analyst / MLView

Adjudicator—

Reg. affairs—

External viewer—

IRB amendment + Safe Harbor

Adjudicated outcomes

PI / co-PIView

CoordinatorView

CT physicist—

BiostatisticianExport

Analyst / MLView

AdjudicatorEdit

Reg. affairs—

External viewer—

Endpoints committee sign-off

Biomarkers / labs

PI / co-PIView

CoordinatorEdit

CT physicist—

BiostatisticianExport

Analyst / MLView

Adjudicator—

Reg. affairs—

External viewer—

Safe Harbor + LIS extract

ECG / echo / PROs

PI / co-PIView

CoordinatorEdit

CT physicist—

BiostatisticianExport

Analyst / MLView

Adjudicator—

Reg. affairs—

External viewer—

Safe Harbor

Physics / QA

PI / co-PIView

Coordinator—

CT physicistEdit

BiostatisticianView

Analyst / MLView

Adjudicator—

Reg. affairs—

External viewer—

Site QA sign-off

Medications

PI / co-PIView

CoordinatorEdit

CT physicist—

BiostatisticianExport

Analyst / MLView

Adjudicator—

Reg. affairs—

External viewer—

Safe Harbor + RxNorm map

Biospecimen inventory

PI / co-PIView

CoordinatorEdit

CT physicist—

Biostatistician—

Analyst / ML—

Adjudicator—

Reg. affairsAudit

External viewer—

Biobank MTA (if enabled)

Data domain	PI / co-PI	Coordinator	CT physicist	Biostatistician	Analyst / ML	Adjudicator	Reg. affairs	External viewer	Governance gate
Identifiers (crosswalk)	Audit	Edit	—	—	—	—	Audit	—	DUA executed + honest broker
Planning CT + DICOM	View	Edit	Edit	View	View	—	—	—	IRB amendment + de-id complete
RT plans / structures / dose	View	Edit	View	View	View	—	—	—	IRB amendment
Clinical & demographics	View	Edit	—	Export	View	—	—	—	IRB amendment + Safe Harbor
Adjudicated outcomes	View	View	—	Export	View	Edit	—	—	Endpoints committee sign-off
Biomarkers / labs	View	Edit	—	Export	View	—	—	—	Safe Harbor + LIS extract
ECG / echo / PROs	View	Edit	—	Export	View	—	—	—	Safe Harbor
Physics / QA	View	—	Edit	View	View	—	—	—	Site QA sign-off
Medications	View	Edit	—	Export	View	—	—	—	Safe Harbor + RxNorm map
Biospecimen inventory	View	Edit	—	—	—	—	Audit	—	Biobank MTA (if enabled)

PI / co-PI

Full oversight of scope, aims, and publication.

Coordinator

Site data extraction, submission, and QA against the master dictionary.

CT physicist

Physics/QA templates, phantom acquisitions, kernel-harmonization outputs.

Biostatistician

Harmonized, de-identified, role-gated analytic tables only.

Analyst / ML

De-identified imaging + tabular data inside the honest-broker enclave.

Adjudicator

Endpoint-only workspace: source documents, imaging, and adjudication log.

Reg. affairs

IRB, DUA, amendment status, and access-log audits — no PHI.

External viewer

Published, aggregate, de-identified summaries only.

Data Model (ERD) & Variable Mapping

Entity-relationship view of how the 11 templates link throughsubject_id,scan_id,plan_id, andevent_id, and how the master dictionary rows map onto each template and each stage of the study data flow.

All templates key back to subject_id. Scans key to plans viascan_id; outcomes key to adjudication viaevent_id; physics/QA keys imaging viascanner_id.

Entity	Template	Primary key	Foreign keys	Flow stage
Subject	02_clinical_demographics.csv	subject_id	—	1. Extraction
Scan	01_imaging.csv	scan_id	subject_id	1. Extraction / 5. Harmonization
RT plan	03_oncologic_treatment.csv	plan_id	subject_id	1. Extraction
Outcome event	04_outcomes_longitudinal.csv	event_id	subject_id	6. Analysis
Biomarker draw	05_biomarkers_labs.csv	subject_id + collection_date	subject_id	1. Extraction
Functional / PRO	06_functional_pro.csv	subject_id + assessment_date	subject_id	1. Extraction
Physics / QA	07_physics_qa.csv	scanner_id + phantom_scan_id	scanner (imaging)	5. Harmonization
Medication	08_medications.csv	med_id	subject_id	1. Extraction
Adjudication	09_adjudication.csv	event_id + round	event_id, subject_id	6. Analysis
Biospecimen	10_biospecimen.csv	specimen_id	subject_id, scan_id, plan_id	1. Extraction
Master dictionary	00_data_dictionary.csv	variable_name	domain, form	4. Central ingest (validation)

Data Management Plan

NIH-aligned Data Management and Sharing plan governing the HIER × Robins Lab collaboration. Applies to all data collected or derived under the umbrella IRB and any executed amendments.

Data types and volume

•Retrospective planning CT DICOM (~1–3 GB per subject after de-identification).
•RT structure sets, plans, and dose grids (~100 MB per plan).
•Derived tabular data across 11 CSV domains at ~2–5 MB per subject.
•Phantom QA acquisitions and calibration logs held by the Robins Lab.
•Optional biospecimen inventory metadata (no material transfer under the pilot).

Standards and interoperability

•DICOM for imaging; DICOM-RT for structures, plans, and doses.
•RxNorm for medications; ICD-10 and ICD-O-3 for diagnoses.
•OMOP-compatible extraction for clinical variables where feasible.
•CDISC-inspired variable naming (snake_case, unit-suffixed).
•FAIR principles: findable, accessible, interoperable, reusable.

De-identification

•HIPAA Safe Harbor for tabular data; expert determination for DICOM.
•Per-subject date shift applied uniformly across all files.
•DICOM header scrubbing (RSNA CTP profile plus a site overlay for private tags).
•Pixel-level burned-in PHI removed by automated OCR and manual QC of at least 5%.
•Face de-identification for any head-inclusive volumes.

Quality assurance

•Ingest validation against 00_data_dictionary.csv (type, range, enum).
•Cross-file referential integrity across subject_id, scan_id, plan_id, event_id.
•Duplicate detection, temporal plausibility, and unit sanity ranges.
•Reviewer double-entry on 10% of adjudicated outcome events.
•Signed SHA-256 manifest per submission batch; audit-logged ingestion.

Storage, access, and security

•Encrypted at rest (AES-256) and in transit (TLS 1.3).
•Role-based access: PI, coordinator, analyst, physicist, viewer (least privilege).
•Institutional identity federation with SSO and MFA; no shared credentials.
•Segregated enclaves per site with a coordinating-center honest-broker workspace.
•Access logs retained per NIH and OHRP guidance; quarterly access reviews.

Retention, sharing, and reuse

•Retain source-linked data for at least 10 years post-publication.
•De-identified derivations shared via controlled-access repository (dbGaP-equivalent) after primary analysis.
•Data Use Agreement required for downstream reuse; no re-identification permitted.
•Code (harmonization pipeline, feature extraction) released under a permissive license on publication.
•Model weights released with a TRIPOD+AI report and a model card.

Data flow and governance

Site extraction. Each site extracts source data from PACS, TPS, EHR, LIS, and tumor registry into the 11 CSV templates plus the DICOM package.
Local de-identification. Site applies Safe Harbor rules, date shift, and DICOM scrubbing before anything leaves the institution.
Manifest and transfer. Files are hashed (SHA-256), listed in a JSON manifest, and pushed to the coordinating-center SFTP endpoint under the DUA.
Central ingest. Coordinating center validates against the master dictionary, resolves referential integrity, and stages data in the honest-broker workspace.
Harmonization. Robins Lab pipeline applies kernel synthesis, virtual monoenergetic mapping, and residual correction; derived biomarkers are regenerated.
Analysis. Statisticians and modelers operate on the harmonized, role-gated dataset; every query is logged.
Publication and release. De-identified derivations and code are released to collaborators; a model card and TRIPOD+AI report accompany any deployed model.

Proposed next steps

Introductory call to confirm scope, co-PI structure, and umbrella IRB scope-check.
Circulate this three-aim brief and data dictionary to the Dartmouth statisticians for feasibility review.
Draft IRB amendment and Data Use Agreement in parallel.
Target Prouty Pilot submission cycle; scope Hitchcock Foundation in parallel.
Stand up a shared, harmonized planning-CT sandbox for reproducibility work.

Contact HIER Institute Prospective protocol

Questions & feedback on this proposal

Reviewers, potential collaborators, statisticians, IRB staff, and Robins Lab members are invited to send questions, corrections, or scoping suggestions on any part of this proposed collaboration. Messages route directly to the HIER Institute leadership team.

NameEmail

AffiliationTopic

Message *0/4000Attachment (optional)PDF, Word, Excel, CSV, images, or text. Max 10 MB.

Privacy notice. Information you submit — including your name, email, affiliation, message, and any attached files — is stored securely in the HIER Institute backend and used only to respond to your inquiry and improve this proposed collaboration. Access is restricted to HIER Institute administrators. Do not submit protected health information (PHI) or other confidential patient data through this form. You may email connect@myheartopulence.com at any time to request deletion of your submission.

Email directly: connect@myheartopulence.com

Future ideas to explore

Adjacent aims from earlier scoping discussions. Retained here for reference; not part of the current locked plan.

Future idea · click to expand

Harmonize quantitative cardiovascular biomarkers across scanners, reconstruction kernels, and dose levels using radiation planning CTs.

HypothesisVendor, kernel, slice thickness, and iterative reconstruction settings introduce systematic bias in coronary artery calcium (CAC), epicardial adipose tissue (EAT), thoracic aortic calcification, and left-ventricular myocardial texture that can be corrected with a physics-informed harmonization pipeline validated against phantom and paired in-vivo scans.

Approach

Phantom-based ground-truth acquisition across Dartmouth and HIER scanners (anthropomorphic thorax phantom, calcium and adipose inserts).
Paired retrospective planning CTs reconstructed with matched and mismatched kernels and thicknesses to quantify bias.
Robins Lab-led CT physics harmonization (kernel synthesis, virtual monoenergetic mapping, ComBat-GAM residual correction).
Re-evaluation of CAC reproducibility, EAT volume and attenuation, and radiomic ICC before and after harmonization.

Deliverables

Open harmonization toolkit (Docker and Python API).
Multi-scanner CT phantom and in-vivo reproducibility dataset (de-identified).

Future idea · click to expand

Derive and externally validate a planning CT cardiovascular risk phenotype (CAC + EAT + aortic calcification + myocardial radiomics) for lifetime MACE in thoracic radiotherapy cohorts.

HypothesisAn integrated multi-structure planning CT phenotype outperforms Agatston CAC alone and conventional clinical risk scores (ASCVD, SCORE2) for predicting MACE in patients undergoing thoracic radiotherapy for breast, lung, esophageal, and lymphoma indications.

Approach

HIER cohort (retrospective planning CTs plus longitudinal cardio-oncology outcomes) as the derivation set.
Dartmouth cohort under the umbrella IRB as the external validation set.
Multitask deep-learning model with radiomic and deep features and a competing-risks time-to-event head (Fine–Gray).
Calibration, decision-curve, and subgroup analyses stratified by sex, age, cancer site, and mean heart dose.

Deliverables

Locked, externally validated model with TRIPOD+AI report.
Nomogram and web calculator integrated into the HIER dashboard.

Future idea · click to expand

Model cardiac substructure dose–response for cardiovascular events using auto-segmented planning CT and harmonized biomarkers, and derive substructure dose constraints.

HypothesisMean and near-max doses to specific cardiac substructures (LAD, left main, LV, LA, sinoatrial node, atrioventricular node) — combined with the harmonized baseline phenotype — reveal dose–response thresholds not visible with whole-heart mean dose.

Approach

Deploy validated cardiac substructure auto-segmentation (TotalSegmentator plus in-house refinement) across both institutions.
Recompute DVH metrics on harmonized planning CTs; extract EQD2-corrected doses.
Cox and machine-learning competing-risks models with baseline phenotype × substructure dose interaction terms.
Bootstrap-derived dose constraints with uncertainty intervals, benchmarked against RTOG and QUANTEC.

Deliverables

Substructure-level dose constraints report.
Integration into HIER treatment planning QA dashboard.
Preliminary data package for NIH R01/P01 or DoD BCRP submission.