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

This study introduces the TRIAD-HFpEF machine learning framework, which integrates multimodal clinical data to probabilistically diagnose Heart Failure with Preserved Ejection Fraction (HFpEF) in the UK Biobank, thereby enabling a 45-fold expansion of genetic discoveries and the identification of 11 actionable therapeutic targets while distinguishing causal drivers from non-causal biomarkers.

The Gist

The Big Problem: The "Invisible" Heart Failure

Imagine Heart Failure with Preserved Ejection Fraction (HFpEF) as a car that has a perfectly working engine (it pumps blood just fine), but the suspension is so stiff and bumpy that the ride is terrible. This condition affects over 30 million people, yet doctors have very few medicines that actually fix the root cause.

Why is it so hard to fix? Because it's like trying to find a needle in a haystack, but the haystack is made of fog. In massive medical databases (like the UK Biobank), there are no clear "HFpEF" labels. The data is messy, incomplete, and often just says "heart problem" without the details. Without clear labels, scientists can't run the genetic tests needed to find new drugs.

The Solution: The "Digital Twin" Detective

The researchers built a super-smart AI detective called TRIAD-HFpEF. Instead of waiting for a doctor to write a perfect label, this AI looks at three different clues to figure out who has the "stiff suspension" heart:

1. The ECG (The Electrical Map): Like listening to the car's engine hum to hear if the rhythm is off.
2. The MRI (The 3D Blueprint): Like taking a high-definition video of the car's chassis to see if the walls are too thick.
3. The Blood Work (The Chemical Smog): Like checking the oil and exhaust fumes for signs of trouble.

The AI was trained on thousands of real patients at Stanford Hospital where the diagnosis was known. Once it learned the patterns, the team sent this "Digital Twin" into the UK Biobank (a massive database of 500,000+ people) to scan everyone.

The Magic Trick: Instead of saying "Yes/No" (which is often wrong), the AI gave everyone a probability score. It's like a weather forecast: "There is a 75% chance of heart failure." This turns a fuzzy, uncertain diagnosis into a clear, measurable number that scientists can use for research.

The Discovery: Finding the Genetic "Smoking Guns"

Once the AI gave everyone a score, the researchers ran a massive genetic search. Before this study, scientists only knew of two genetic spots linked to this heart failure.

Thanks to the AI's new way of measuring the disease, they found over 90 new genetic spots! That's a 45-fold increase in knowledge. It's like going from having a map with two cities marked to a map showing the entire highway system.

They found that the disease is driven by three main things:

• Metabolism: How the body handles energy and weight (linked to the famous FTO gene).
• Development: How the heart was built in the first place.
• Inflammation: The body's internal fire alarms going off.

The Gold: Therapeutic Targets vs. Red Herrings

The most exciting part is how they used this data to find new medicines. They looked at thousands of proteins in the blood and asked two questions:

1. Is this protein causing the problem? (If we block it, will the patient get better?)
2. Is this protein just a symptom? (If we block it, will it just be like putting a bandage on a broken leg?)

The Winner: FLT3 (The Guardian)

The AI pointed to a protein called FLT3. The data suggested that having more FLT3 protects the heart, while having less hurts it.

To prove this, the researchers looked at leukemia patients who were taking drugs to block FLT3 (to kill cancer). They found that these patients developed heart failure with the exact same "stiff suspension" symptoms.

• The Analogy: Imagine you have a shield that protects your heart. The leukemia drugs accidentally took the shield away. The heart didn't break because of the cancer; it broke because the shield was gone. This proves that boosting FLT3 could be a new way to treat heart failure.

The Loser: MPO (The Smoke, Not the Fire)

Another protein, MPO, was a huge red flag in previous studies. People thought blocking it would cure heart failure. But the AI showed that MPO levels only go up after the heart is already damaged.

• The Analogy: MPO is like the smoke coming from a house fire. If you spray water on the smoke (block MPO), the fire (the heart failure) keeps burning. Recent clinical trials tried to block MPO and failed. This study explains why: MPO is just a symptom, not the cause.

The Bottom Line

This paper is a game-changer because it used AI to clean up the messy data, allowing scientists to finally see the genetic causes of a disease that was previously invisible to them.

• They found 90+ new genetic clues.
• They identified a promising new drug target (FLT3) that needs to be tested.
• They saved the medical community from wasting time and money on a dead-end target (MPO).

It's a perfect example of how combining computer science (AI) with biology can solve medical mysteries that have stumped doctors for decades.

Technical Summary

1. Problem Statement

Heart Failure with Preserved Ejection Fraction (HFpEF) affects over 30 million people globally but lacks effective disease-modifying therapies. A major bottleneck in discovering new treatments is the imprecise phenotyping of HFpEF in large-scale biobanks (e.g., UK Biobank).

• The Challenge: HFpEF diagnosis is inherently probabilistic and requires echocardiographic evidence of diastolic dysfunction, which is often missing in biobank datasets. Consequently, biobanks rely on non-specific ICD codes (e.g., "congestive heart failure"), leading to noisy data.
• The Consequence: Previous Genome-Wide Association Studies (GWAS) using these imprecise definitions have identified only two marginally significant loci ($FTO$ and $IGFBP7$), severely limiting genomic discovery and drug target identification.

2. Methodology: The TRIAD-HFpEF Framework

The authors introduced TRIAD-HFpEF (Tri-modal Assessment and Discovery of HFpEF), a machine learning framework designed to generate continuous, high-fidelity HFpEF phenotypes from multimodal data. The study followed a four-phase pipeline:

A. Phenotype Construction (Training Data)

• Source: Real-world clinical data from Stanford Health Care ($N=115,149$ patients).
• Definition: HFpEF cases were defined via EHR codes or, more granularly, by combining:
  • Preserved LVEF (>50%).
  • Abnormal diastolic function (extracted via regex from echo reports: $E/e' > 14$, $e' < 7/10$, TR velocity $> 2.8$, LA volume index $> 34$).
  • Elevated natriuretic peptides (NT-proBNP/BNP) and diuretic use.
  • Exclusion of congenital heart disease, transplants, and pulmonary hypertension.
• Granularity: The $H2FPEF$ score was also calculated to create a continuous probability phenotype.

B. Model Training (Three Modalities)

Three separate deep learning models were trained to predict HFpEF probability:

1. TRIAD-ECG: A deep convolutional neural network (CNN) with channel and spatial attention mechanisms processing raw 12-lead ECGs (5000 data points). It was combined with 117 numeric ECG features via XGBoost.
   • Test AUC: 0.81.
2. TRIAD-CMR: A ResNet V2 model adapted to process 24 images (3 views $\times$ 8 frames) from Cardiac Magnetic Resonance (CMR) cine loops.
   • Test AUC: 0.74.
3. TRIAD-LAB: An XGBoost model using 39 clinical covariates (demographics, labs, medication history).
   • Test AUC: 0.84.

C. Deployment and Validation (UK Biobank)

• The trained models were deployed on the UK Biobank (UKB) to assign a continuous HFpEF probability score to participants ($N \approx 502k$ for LAB, $N \approx 81k$ for ECG, $N \approx 84k$ for CMR).
• Validation: Since UKB lacks ground-truth HFpEF labels, validation was performed via:
  • Structural/Functional Correlation: Strong associations with expected HFpEF features (e.g., increased Left Atrial volume, increased myocardial mass, preserved LVEF).
  • Specificity: Correctly classifying patients with LVEF < 50% as non-HFpEF (87% accuracy for LAB).
  • Outcomes: Significant prediction of HF hospitalization, non-specific HF ICD codes, and cardiovascular mortality.

D. Multi-Omic Analysis

The predicted probabilities served as quantitative traits for:

• GWAS: Conducted separately for each modality.
• Proteomics: Proteome-wide association studies (PWAS) using UKB-Pharma Proteomics Project data.
• Causal Inference: Bidirectional Mendelian Randomization (MR) and Colocalization to distinguish causal therapeutic targets (Protein $\to$ HFpEF) from disease consequences/biomarkers (HFpEF $\to$ Protein).

3. Key Results

A. Genomic Discovery

• Expansion of Loci: The study identified 101 independent genome-wide significant loci (P < $5 \times 10^{-8}$), a 45-fold increase over previous studies.
• Convergence: The $FTO$ locus was significant across all three modalities, highlighting a central metabolic axis.
• Modality-Specific Signals:
  • ECG: Enriched for cardiac development and electrophysiology genes (e.g., $SOX5$, $CCDC141$).
  • CMR: Highlighted inflammatory pathways (e.g., $SAA2-SAA4$).
  • LAB: Strong metabolic and inflammatory signals (e.g., $PRKAG2$, $ITGA4$).
• Heritability: Estimated HFpEF heritability at ~9%.

B. Proteomic and Causal Prioritization

Using a tiered evidence framework (PWAS + Colocalization + Forward MR + Reverse MR exclusion), the authors prioritized:

1. 11 Therapeutic Targets (Causal): Proteins where higher expression is protective or causal.
   • Key Finding: FLT3. Genetically predicted higher FLT3 expression is protective.
   • Key Finding: PRKAG2. Implicates AMPK signaling and glycogen storage pathways.
2. 7 Biomarkers (Consequences): Proteins altered after disease onset (Reverse MR significant, Forward MR not).
   • Key Finding: MPO (Myeloperoxidase). Consistent with recent negative clinical trials for MPO inhibitors.

C. Clinical Validation of FLT3

• Hypothesis: If FLT3 is protective, FLT3 inhibitors (used in leukemia) should worsen diastolic function.
• Validation: In an independent Stanford cohort of leukemia patients treated with FLT3 inhibitors:
  • Result: Significant worsening of diastolic function (decreased $e'$, increased $E/e'$) post-treatment.
  • Sparing: No significant change in LVEF (systolic function) or TR velocity.
  • Conclusion: This confirms FLT3 inhibition specifically impairs diastolic function, validating the causal link.

4. Significance and Impact

• Methodological Breakthrough: Demonstrates that "digital twin" approaches using multimodal ML can overcome the lack of binary diagnostic codes in biobanks, transforming noisy data into quantitative phenotypes for high-power genetic discovery.
• Genomic Scale: The largest HFpEF GWAS to date, expanding known genetic architecture from 2 to 101 loci.
• Therapeutic Prioritization:
  • FLT3: Identified as a novel cardioprotective target. The study provides orthogonal validation (genetic, preclinical, RCT data, and real-world clinical validation) suggesting FLT3 activation could be a therapeutic strategy for HFpEF.
  • De-prioritization: Successfully identified MPO as a downstream biomarker rather than a causal target, explaining the failure of MPO inhibitor trials (ENDEAVOR, SATELLITE).
• Open Science: All model code, weights, and molecular findings are open-sourced to accelerate future research.

5. Limitations

• Lack of Ground Truth: Validation in UKB relied on proxy measures (structural features, outcomes) rather than confirmed HFpEF diagnoses, though the authors argue continuous probabilities are superior to binary codes.
• Population: Primarily European ancestry (UKB), limiting generalizability to other ethnic groups.
• Model Calibration: While relative ranking is preserved for GWAS, absolute probability calibration may vary.

Conclusion

TRIAD-HFpEF represents a paradigm shift in complex disease research, moving from static, imprecise phenotyping to dynamic, multimodal probabilistic modeling. By integrating ECG, imaging, and biomarkers, the framework successfully unlocked the genomic and proteomic architecture of HFpEF, identifying actionable therapeutic targets (like FLT3) while filtering out non-causal associations.