A graph neural network (GNN) framework integrating plasma proteomics, clinical variables, and molecular prior-knowledge data achieved an external test AUC of 0.82 for predicting heart failure (HF) following acute myocardial infarction (MI) — outperforming XGBoost and generalized linear models (best AUC 0.77) across two independent public proteomic cohorts. Attention-mechanism GNNs drove the performance edge. Regression models applied to longitudinal data from over 400 patients in the EMMY empagliflozin trial predicted biomarker treatment-response changes with an RMSE of 0.56. Feature importance consistently flagged NT-proBNP, cardiac troponins TNNI3/TNNT2, and prior HF history as top predictors.
Post-MI heart failure affects roughly 25–40% of survivors and carries grim prognosis, yet clinicians still rely on a narrow biomarker set with modest stratification power. By encoding each patient as a personalized knowledge graph, this framework exploits protein–protein interaction topology alongside traditional clinical features — a systems-biology leap beyond tabular ML. The AUC improvement from 0.77 to 0.82 using biomarkers alone is clinically meaningful if it translates to real-world triage. That said, important caveats apply: datasets are retrospective and publicly sourced, cohort demographics and assay platforms may limit generalizability, and the EMMY regression analysis is exploratory. The finding that graph architecture outperforms ensembles like XGBoost is notable but not yet paradigm-shifting — similar GNN advantages have been reported in oncology risk models. Critically, this is a preprint posted on medRxiv and has not yet undergone peer review; performance metrics and mechanistic conclusions should be treated as preliminary until independent validation is complete.
