Abstract
Cardiovascular disease is one of the leading causes of death across the world, indicating the clinical limits of existing early detection and diagnosis approaches. The traditional machine learning models used for heart disease detection heavily depend on centralised databases, which pose a significant risk of violating data privacy and confidentiality. In order to overcome these limitations, Federated Learning (FL), a decentralized and privacy-preserving paradigm for training models in collaboration between many devices while keeping raw patient data at the device level, has been proposed. We present in this systematic exploratory review an overview of the most FL approaches and methods that have been used with respect to heart disease detection. In this work we review widely used evaluation metrics and discuss major issues (heterogeneity in data distributions, high communication overhead and the possible derailing of model accuracy). In addition, the limitations of several state-of-the-art solutions are presented, such as data partitioning strategies, federated aggregation techniques and advanced encryption mechanisms. The results show that although FL is more powerful than the DIR methods to protect data in medical applications, FL either takes time to converge or does not converge when faced with non-IID and heterogeneous clinical data. Over the practical studies presented here, the promise of FL both improving the prediction accuracy is corroborated, while able to secure sensitive information. The paper also presents future directions by highlighting the adoption of FL with latest technologies including IoMT and Blockchain for better scalability and security. Lastly, we highlight specific research opportunities, such as learning aggregation and optimization algorithms, non-IID data approaches and real-time federated learning clinical platforms.
Keywords: Federated Learning, Healthcare Diagnostics, Heart Disease Detection, Machine Learning, Privacy- preserving.