F1000Res. 2026 Jan 31;14:1341. doi: 10.12688/f1000research.171338.2. eCollection 2025.
ABSTRACT
BACKGROUND: Myocardial ischemia is a dynamic, complex process characterized by hyperkalemia, acidosis, and ATP depletion. While these three conditions alter cardiomyocyte electrophysiology, it is difficult to discern how much each one individually contributes to the resulting changes in action potential (AP). In this study, we test whether machine learning can deconvolute these distinct ischemic patterns within a single AP.
METHODS: We developed a multi-target regression model trained on data generated by the Luo-Rudy (1991) computational model of a ventricular cardiomyocyte, simulating a wide range of ischemic conditions. The model was designed to predict two continuous variables: extracellular potassium concentration ([K +]o) and intracellular pH (pHi).
RESULTS: The model achieved high accuracy on a held-out test set, with mean squared errors (MSE) below 0.25 for [K +]o and below 0.01 for pHi. To further generalize this model, we applied this trained model to a structurally distinct model, the Ten Tusscher (2006) framework. We were able to accurately predict [K +]o and pHi from APs, demonstrating that the learned principles are robust. A feature importance analysis revealed that resting membrane potential (RMP) was the strongest predictor for [K +]o, while action potential duration (APD) is most important for predicting pHi, underscoring these distinct cardiomyocyte electrophysiological patterns.
CONCLUSIONS: Our approach can distinguish distinct ischemic drivers and has potential for in silico drug screening and mechanistic analysis.
PMID:41685048 | PMC:PMC12891950 | DOI:10.12688/f1000research.171338.2