FEDERATED LEARNING-BASED ATRIAL FIBRILLATION DETECTION USING DEEP LEARNING AND GENERATIVE AI-BASED RISK EXPLANATION

Abstract

Cardiovascular diseases, which in short terms are called “CVD,” are linked to a group of diseases related to the heart, brain, and many disorders of blood vessels. Atrial Fibrillation is also a common type of CVD which can cause heart disorders and brain strokes. We research the early detection of this dangerous heart disease so that we can save the lives of patients by spreading knowledge about this irregular heartbeat. Our research uses a technique which can protect the privacy of patients’ data because there are also risks of data leakage, and we built a combined CNN and BiLSTM model to reduce this leakage of data. We use the Federated Learning technique in which we can solve the issue of privacy and data security of patients. In this technique, we also use 12-Lead ECG data for binary AF classification, and we use the Federated Averaging (FedAvg) method to train our decentralized model across three local nodes. But we have imbalanced data inside each node, and we solve this issue by using the Synthetic Minority Oversampling Technique (SMOTE). We set up an edge-based pipeline that runs a local Phi-3-mini-4k-instruct language model directly on the device to explain the potential risks of stroke, irregular heartbeat complications, and heart failure after the detection of Atrial Fibrillation. This Generative AI system strictly follows all privacy rules of General Data Protection Regulation (GDPR) and Health Insurance Portability and Accountability Act (HIPAA) to create medical reports and understandable summaries for doctors in hospitals, which is a good point of this system. When we trained our model for AF prediction, our federated model achieved 94% accuracy, which is greater as compared to the centralized model in which we achieved 91% accuracy. This comparison of accuracy shows that we can manage data privacy with AF prediction. In the end, our project proves that we can safely screen for AF in the real world.