Deconvoluting the Electrophysiological Signatures of Myocardial Ischemia using a Validated Machine Learning Framework

Author summary

During a myocardial infarction, or heart attack, cardiac myocytes experience an oxygen deficiency, increase in potassium outside the cell, and an increase in intracellular acid. These conditions contribute to an altered AP, however, it is difficult to discern the individual contributions of each condition. In this study, we explored whether a computer could learn to identify distinct signatures of damage. By simulating ischemia in a virtual heart cell, we generated thousands of APs using trusted computer models. With a trained machine learning system with high accuracy, we then analyzed the shape of these APs to predict the levels of extracellular potassium and intracellular acidity. To ensure the generalizability of our findings to further models, we tested our method on an entirely different, modern cardiac cell model. Our system maintained high accuracy despite its lack of familiarity with this model, supporting the robustness of our approach. Finally, we showed that the system could be used to evaluate how well a simulated drug protects heart cells, suggesting a new direction for testing therapies virtually.

Introduction

Myocardial ischemia, a defining feature of heart attacks, initiates a series of interconnected pathological processes at the cellular level. As blood flow becomes restricted, oxygen supply rapidly diminishes, inhibiting aerobic metabolism and depleting ATP. This energy shortage leads to a cascade of disruptions in the cardiomyocyte’s internal environment, most notably, a rise in extracellular potassium concentration ([K⁺]o) due to Na⁺/K⁺ pump failure, and a drop in intracellular pH (pHi) caused by lactate buildup through anaerobic glycolysis.²

Together, these stressors, hyperkalemia, acidosis, and ATP depletion, profoundly alter the electrophysiological properties of cardiomyocytes. Hyperkalemia causes depolarization of the RMP, while acidosis interferes with the function of key ion channels, including the fast sodium current (INa) and the L-type calcium current (ICaL). At the same time, ATP depletion activates the ATP-sensitive potassium current (IK (ATP)), which significantly shortens the APD.³ The resulting ischemic AP reflects a complex interplay of these pathophysiological changes.

Although the general shape of ischemic APs is well documented, characterized by depolarized RMP, slower upstroke, and shortened duration, it remains extremely difficult to determine which specific pathological factor is primarily responsible for these changes in a given case. For instance, two cardiomyocytes may display similarly short APDs due to different underlying combinations of hyperkalemia and acidosis. The ability to isolate and quantify the individual effects of these contributors could greatly improve our understanding of ischemic tissue states and support the development of more precisely targeted therapies.

In this study, we explore the potential of a machine learning approach, specifically a multi-target regression framework, to infer the distinct influences of hyperkalemia and acidosis from AP waveform features. We propose a three-phase strategy to investigate this hypothesis. First, we train Random Forest regressors on a synthetic dataset generated using the Luo-Rudy (1991) computational model. Next, we test the generalizability of the trained models using an independent dataset created with the structurally distinct Ten Tusscher (2006) model. Finally, we apply the validated framework to a simulated drug intervention in a severely ischemic cell to evaluate its potential use in quantifying therapeutic effects in silico.

Materials and methods

Results

Discussion

This study introduces a machine learning-based method for inferring specific ischemic stressors, hyperkalemia and acidosis, from single AP waveforms. Our results show that although these conditions co-occur and influence similar AP features, their electrophysiological “signatures” can be analyzed and separated by trained models. That said, the relatively modest dataset size (150 training APs and 50 validation APs) raises questions about robustness. Although the regressors performed well, expanding the training set or applying data augmentation strategies would strengthen confidence in model generalizability and reduce the risk of overfitting.

Crucially, the regressors trained on the Luo-Rudy model maintained high predictive accuracy when applied to the Ten Tusscher model. We selected Random Forest regressors for their interpretability and robustness with limited data. Nonetheless, alternative machine learning methods, such as gradient boosting or neural networks, could capture nonlinearities differently. Future work may compare performance across these approaches to identify optimal strategies for ischemic feature deconvolution. This cross-model validation provides strong evidence that the learned relationships reflect underlying principles of cardiac electrophysiology.

Feature importance analysis reinforced physiological expectations: RMP reflected extracellular potassium levels,² while APD and amplitude were indicative of pH-related effects.⁶ The model’s successful use in simulating and quantifying a drug’s electrophysiological impact also points to its utility as a screening tool in computational pharmacology.

Limitations include the use of simplified cell models and the exclusion of other ischemic contributors like mechanical stretch, sympathetic stimulation, or reactive oxygen species. While our framework demonstrates strong predictive accuracy in silico, it has not yet been tested against experimental data. A clear next step is validation against electrophysiological recordings in ischemic cardiomyocytes, either through patch-clamp experiments or extracellular field potential measurements in animal ischemia models. Such experimental validation would not only strengthen biological plausibility but also provide critical insight into the translational robustness of the approach.

Additionally, although ATP-sensitive currents were incorporated through IK (ATP), the framework did not directly predict ATP depletion as an output. Extending the model to infer ATP levels would provide a more complete representation of ischemic pathophysiology, particularly in the context of metabolic stress. Future work should also explore tissue-level simulations and intercellular coupling to bridge the gap between single-cell models and translational applications.

Nonetheless, the present findings offer a strong foundation for data-driven approaches to studying and treating myocardial ischemia. In particular, future applications could move beyond single-cell models to incorporate tissue-level simulations, extracellular field potentials, or even clinical ECG recordings. Such extensions would help bridge the gap between mechanistic insight and translational utility in human patients.

Ethics and consent

Ethical approval and consent were not required for this study.