Prediction of Myocardial Infarction Based on Non-ECG Sleep Data Combined With Domain Knowledge

Abstract

Prediction of myocardial infarction (MI) is crucial for early intervention and treatment. Machine learning has increasingly been applied in the realm of disease prediction. This study explores the feasibility of utilizing easily obtainable heart rate (HR) and respiratory rate (RR) data collected during nocturnal sleep, in conjunction with clinical characteristics and medical domain knowledge, to predict MI. Data for this investigation were sourced from the Sleep Heart Health Study (SHHS) program in the United States, which was categorized into MI and non-MI groups based on the occurrence or absence of MI during follow-up, involving a total of 488 participants. Multiple features related to HR and RR were extracted and integrated with clinical features; four algorithms—MLP, SVM, XGBoost, and CNN—were employed for model construction. The findings indicated that the MLP model exhibited superior performance, achieving an accuracy rate 71.1%. Furthermore, three medical rules age, HR, and RR were incorporated into the MLP model to mitigate the limitations of small sample sizes. The experiments demonstrate that the model’s accuracy reaches its optimal level by combining the age rule, improving to 73.1%. The findings indicate that leveraging non-cardiac electrophysiological data obtained during sleep alongside medical domain knowledge can significantly enhance the accuracy of early predictions regarding cardiac MI while offering novel insights for its prevention and diagnosis.