A multimodal machine learning system called TRIAD-HFpEF identified over 90 novel genetic loci associated with heart failure with preserved ejection fraction (HFpEF), representing a 45-fold increase from the two previously known variants. The framework integrated electrocardiograms, cardiac MRI, and biomarkers to assign HFpEF probabilities in UK Biobank participants, distinguishing 11 causal therapeutic targets from 7 non-causal biomarkers. This represents a significant breakthrough for a condition affecting over 30 million people worldwide that currently lacks disease-modifying treatments. The AI approach overcomes traditional limitations in HFpEF research where biobanks lack specific diagnostic codes and echocardiograms, instead treating diagnosis as probabilistic rather than binary. Notably, the system identified FLT3 as a therapeutic target, validated by clinical data showing diastolic function deterioration with FLT3 inhibitors, and correctly flagged MPO as non-causal, consistent with recent negative clinical trials. However, this preprint awaits peer review, and the machine learning predictions require validation in diverse populations. The methodology could transform genetic discovery for other complex cardiovascular syndromes where precise phenotyping remains challenging.