Teaching machines to diagnose what humans can’t see.

Approximately half of patients referred for cardiac stress testing have impaired coronary vasomotor function despite normal perfusion imaging.

Standard cardiac testing detects flow-limiting epicardial stenosis. Most of the coronary circulation, however, consists of vessels below the resolution of angiography, and impaired vasodilator capacity in these vessels carries prognostic significance independent of anatomic disease.

We quantified coronary flow reserve by cardiac PET in 2,783 patients referred for rest/stress imaging. Patients in the lowest tertile of flow reserve had a 5.6-fold higher risk of cardiac death (95% CI 2.5–12.4) after adjustment for clinical risk factors, left ventricular ejection fraction, and the extent of scar and ischemia. Incorporation of flow reserve into risk models correctly reclassified 34.8% of intermediate-risk patients.

We subsequently examined 1,218 patients without known coronary disease and with visually normal stress perfusion imaging. Coronary microvascular dysfunction, defined as flow reserve below 2.0, was present in 51% of men and 54% of women (equivalence P=0.0002). Flow reserve was similarly predictive of major adverse cardiac events in both sexes, with no significant sex-by-flow-reserve interaction. Among the subgroup with coronary artery calcium scores of zero, microvascular dysfunction remained common (44% of men, 48% of women).

These findings established coronary microvascular dysfunction as a prevalent, prognostically significant condition affecting both sexes, rather than a diagnosis of exclusion specific to women. Current work addresses the reproducibility of flow quantification across centers and scanner platforms, and the therapeutic implications of impaired vasomotor function.

We apply metabolomic, proteomic, and genetic profiling in longitudinal population cohorts to identify molecular signatures that precede clinical cardiovascular disease and to generate hypotheses about the pathways involved.

Conventional risk factors classify populations but provide limited insight into individual pathophysiology, particularly in young adults, in whom event rates are low and risk scores perform poorly. Deep molecular phenotyping in cohorts with extended follow-up offers a route to both earlier risk discrimination and mechanistic inference.

In the CARDIA study, comprehensive metabolic phenotyping in young adulthood refined cardiovascular risk estimation beyond established risk factors, and circulating metabolite profiles in early adulthood identified individuals who later developed diabetes.

More recently, we characterized the plasma proteomic correlates of lipoprotein(a). Lp(a) concentration is largely genetically determined and relatively stable across the lifespan, and although it is an established risk factor for atherosclerotic cardiovascular disease, the pathways linking it to clinical events remain incompletely defined. We measured 184 cardiovascular proteins in 3,920 CARDIA participants without prevalent coronary heart disease and followed them for a median of 27 years. An Lp(a)-associated proteomic signature derived by LASSO regression was independently associated with incident coronary artery calcification (standardized β=0.40, P<0.0001) and hs-CRP (β=0.11, P=0.00015) after adjustment for Lp(a) concentration, whereas Lp(a) concentration itself did not retain significance in the same models. The signature replicated in 37,996 UK Biobank participants. Pathway enrichment implicated interleukin signaling, cell-surface interactions at the vascular wall, and fibrin clot formation.

The cardiac tests with the greatest diagnostic utility are also the least accessible. We develop machine learning methods that extend quantitative capability to widely available modalities.

Chest pain accounts for roughly 8 million emergency department visits and 6 million outpatient visits annually in the United States. Quantitative PET myocardial perfusion imaging provides accurate assessment of myocardial flow reserve and coronary microvascular dysfunction, but is expensive and unevenly distributed. The limiting factor for training AI models on these endpoints is the scarcity of high-quality labels, since labels derive from the same scarce tests.

We addressed this with a self-supervised foundation model. A modified vision transformer was pretrained on 800,035 unlabeled ECG waveforms from MIMIC-IV-ECG, then fine-tuned on smaller labeled datasets derived from PET (N=3,126) and clinical reports (N=13,704) across 12 prediction tasks spanning myocardial perfusion, ventricular function, and rhythm. Self-supervised pretraining improved diagnostic accuracy in 11 of 12 tasks relative to de novo supervised training, with AUROC ranging from 0.763 for impaired flow reserve to 0.955 for reduced ejection fraction. Performance was retained across five independent databases, including UK Biobank and PTB-XL. Label efficiency improved substantially: the pretrained model matched the peak performance of the de novo model using 10% of available training data for the flow reserve task.

A parallel effort addresses attenuation correction in SPECT. Approximately 75% of SPECT myocardial perfusion imaging worldwide is performed without CT attenuation correction. We trained a convolutional network on 11,532 paired studies to generate attenuation-corrected polar maps from non-corrected data. Correlation with CT-corrected reference improved from R²=0.68 to 0.85, and specificity for obstructive disease improved by 18.9% at 88% sensitivity, without additional imaging or radiation exposure.