Heart Attack Prediction Using Federated Learning on Distributed Medical Data

Abstract

Despite advances in cardiology, heart disease continues to be a major global challenge. The development of tools for early detection and accurate prediction of the probability of triggering a heart attack as a critical event in heart disease is essential. Traditional machine learning models for heart attack prediction are a violation of medical data privacy and security as they involve a centralized dataset. Another model, federated learning, is the optimal way to keep decasualized privacy data from being available in multiple medical institutions. In this work, we conduct a study to determine how effective FL is in predicting heart attack using Logistic Regression and Support Vector Machine models with large-scale simulated distributed medical data. The first model yielded an accuracy of 88.52%, indicating that to some extent, heart attack prediction is a use case for FL. We also conduct further research on other models, and the SVM model demonstrated an accuracy of 86.89%, which is considered a fully dependent variable to be predicted as favorable. The current research also examines additional models, including K-Nearest Neighbors and Decision Tree. The latter showed lower performance, exercising an accuracy of 68.89%, while it has higher value in interpretability. It deserves to be aware that the research focus is the communication overhead within the FL framework. In my opinion, it is significant to proceed with the further investigations on the enhancements of optimum communication approaches balancing the model accuracy, training time, and communication cost. Moreover, privacy preservation within the FL deserves to be highlighted. It is worth mentioning that current research is the initial attempt, whereas privacy-preserving techniques customized for LR and SVM within the FL remain an unknown field to be analyzed. Overall, through this research, we have showed the significant potential of the FL approach for heart attack prediction with the use of distributed medical data. This future was proposed by considering the observance of privacy limitations on the accessed datasets. The FL could remain as a significant solution in the development of appropriate machine learning models, enhancing the efficiency of communication, and providing privacy considerations with an opportunity to minimize the risks of compromise.