Multimodal Machine Learning Reveals the Genomic and Proteomic Architecture of Heart Failure with Preserved Ejection Fraction

J O'sullivan, Taedong Yun, Ruoyi Cai, David Amar, Themistocles L Assimes, Nigam H. Shah, Dan Say Kim, E D Lewis, Naresh Khurana, Farhad Hormozdiari, J. Weston Hughes, Gabriel N. Mannis, Michael Salerno, Mark E. Pepin, Navid A. Nafissi, Jared Wallace, Richa Tiwari, Manuel A. Rivas, Charles Lee, Cory Y. McLean, E AshleymedRxiv 2026 · DOI

TL;DR

Researchers developed TRIAD-HFpEF, a multimodal machine learning framework that predicts the probability of Heart Failure with Preserved Ejection Fraction (HFpEF) using ECGs, MRI images, and standard blood markers. By applying these "digital twin" phenotypes to the UK Biobank, they expanded the known genetic landscape of the disease from just 2 loci to 101, identifying 11 promising therapeutic targets and 7 biomarkers that are consequences—rather than causes—of the disease.

What Problem Does This Paper Address?

Heart failure with preserved ejection fraction (HFpEF) is a "stiff heart" condition affecting over 30 million people globally. Unlike other forms of heart failure, it lacks effective disease-modifying treatments.

The bottleneck isn't a lack of data, but a lack of labels. Large-scale genetic repositories like the UK Biobank (UKB) are "gold mines" for discovery, but they lack the specific echocardiograms and diagnostic codes required to identify HFpEF patients. Because the disease exists on a clinical continuum, binary "yes/no" classifications often fail. This paper tackles the "imprecise phenotyping" problem by using machine learning to turn noisy, multimodal clinical data into a precise, continuous probability of disease.

Methods

The authors built TRIAD-HFpEF (TRI-modal Assessment and Discovery of HFpEF) using a four-phase approach:

  1. Phenotype Construction: They used "ground truth" data from Stanford Health Care, employing regular expressions (regex) to extract diastolic function measurements from thousands of physician reports to define high-fidelity HFpEF cases.
  2. Multimodal Training: They trained three distinct models:
  3. TRIAD-ECG: A deep learning model (CNN with attention mechanisms) using raw 12-lead ECG waveforms.
  4. TRIAD-CMR: A ResNet V2 architecture adapted to process 3D+time "cines" (videos) from Cardiac Magnetic Resonance imaging.
  5. TRIAD-LAB: An XGBoost model using 39 standard clinical covariates (age, BMI, blood counts, etc.).
  6. Deployment: These models were deployed to the UK Biobank to assign HFpEF probabilities to over 500,000 participants.
  7. Omic Discovery: They used these predicted probabilities as continuous traits for Genome-Wide Association Studies (GWAS) and proteomic analyses (Mendelian Randomization and colocalization).

Datasets

  • Stanford Health Care Dataset: Used for training and internal validation (n=115,149 for ECG/Labs; n=6,092 for CMR).
  • UK Biobank (UKB): The primary deployment and discovery cohort (n ≈ 500,000).
  • Million Veterans Program (MVP): Used to cross-validate genetic correlations.
  • Independent Clinical Cohort: 104 patients treated with FLT3 inhibitors at Stanford were used to validate one of the top therapeutic targets.

Key Insights and Ideas

  • From Binary to Probabilistic: The core "aha moment" is that treating HFpEF as a continuous probability—a "digital twin" approach—is far more statistically powerful for genetic discovery than using traditional, often inaccurate, diagnostic codes.
  • Causal vs. Correlational: By using bidirectional Mendelian Randomization, the team could distinguish between proteins that cause the disease (targets) and proteins that are simply elevated because of the disease (biomarkers).
  • The FLT3 Connection: They identified FLT3 (a protein often targeted in leukemia) as a protective factor. This explains why cancer patients taking FLT3 inhibitors often develop heart failure—the drug is accidentally "switching off" a cardioprotective mechanism.
  • Debunking MPO: While MPO (myeloperoxidase) is a popular biomarker for HFpEF, this study categorized it as "non-causal." This aligns with recent clinical trials where MPO inhibitors failed to help patients, suggesting MPO is a smoke signal, not the fire.

Key Results

  • Massive Genetic Expansion: The study identified 101 independent genome-wide significant loci, a 45-fold increase over the 2 loci previously known.
  • High Predictive Accuracy: The models achieved strong Area Under the Curve (AUC) scores on held-out test sets: 0.84 for TRIAD-LAB, 0.81 for TRIAD-ECG, and 0.74 for TRIAD-CMR.
  • Validation of Targets: They prioritized 11 therapeutic targets (including FLT3 and PRKAG2) and 7 non-causal biomarkers.
  • Clinical Proof: In their validation cohort, patients treated with FLT3 inhibitors showed a statistically significant worsening of diastolic function (the hallmark of HFpEF) while systolic function remained unchanged, directly confirming the ML-derived hypothesis.

Limitations and Future Work

  • Ground Truth in Biobanks: There is no absolute "ground truth" for HFpEF in the UK Biobank to perfectly verify the predictions; the authors mitigated this by validating against mortality and hospitalization outcomes.
  • Calibration: While the ranking of participants by probability is robust for genetic discovery, the absolute probability values (e.g., 0.75 vs 0.76) may not be perfectly calibrated for individual clinical bedside use yet.
  • Future Directions: The authors have open-sourced their code and weights, inviting the community to apply this multimodal framework to other "complex syndromes" that suffer from imprecise labeling in large datasets.

Summary updated

Abstract

Abstract Heart failure with preserved ejection fraction (HFpEF) affects over 30 million people and lacks disease-modifying therapies. Although genomic-led drug discovery increases success by more than 2.6-fold, HFpEF genomic discovery remains constrained by imprecise phenotyping in biobanks, with only two loci identified to date. Biobanks lack HFpEF diagnostic codes and echocardiograms, yet HFpEF diagnosis exists along a continuum and is inherently probabilistic, presenting an opportunity for multimodal prediction. Here we introduce TRI-modal Assessment and Discovery of HFpEF (TRIAD-HFpEF), a machine learning framework integrating electrocardiograms, cardiac magnetic resonance imaging, and biomarkers to assign HFpEF probabilities. Deployed in UK Biobank, these probabilities validate with respect to mortality, hospitalizations, and structural and functional HFpEF features. Genome-wide and proteomic analyses reveal over 90 novel loci, a 45-fold expansion, and distinguish causal proteins from non-causal biomarkers of disease progression, prioritizing 11 therapeutic targets and 7 non-causal biomarkers. We identify FLT3 as one of the 11 therapeutic targets, consistent with the reported 7-fold increased heart failure risk from FLT3 inhibitors in leukemia. We validate this finding by demonstrating significant worsening of diastolic function following FLT3 inhibitor treatment in an independent clinical cohort. Conversely, MPO emerged as one of the 7 non-causal biomarkers, aligning with three recent negative MPO inhibitor trials. TRIAD-HFpEF demonstrates that machine learning-derived phenotypes can unlock genetic discovery in complex syndromes, identifying actionable targets while deprioritizing associations reflecting disease consequences rather than causes.
BibTeX
@misc{osullivan2026multimodal,
  author = {J O'sullivan and Taedong Yun and Ruoyi Cai and David Amar and Themistocles L Assimes and Nigam H. Shah and Dan Say Kim and E D Lewis and Naresh Khurana and Farhad Hormozdiari and J. Weston Hughes and Gabriel N. Mannis and Michael Salerno and Mark E. Pepin and Navid A. Nafissi and Jared Wallace and Richa Tiwari and Manuel A. Rivas and Charles Lee and Cory Y. McLean and E Ashley},
  title = {{Multimodal Machine Learning Reveals the Genomic and Proteomic Architecture of Heart Failure with Preserved Ejection Fraction}},
  year = {2026},
  doi = {10.64898/2026.02.07.26345811},
  howpublished = {medRxiv},
}
Figure