Hybrid Neural Reconstruction of ECG from PPG for Differentially Private Atrial Fibrillation Detection

This repository contains the implementation and experimental pipeline supporting the research work: "Privacy-Preserving ECG Reconstruction and Atrial Fibrillation Detection from PPG Signals Using Hybrid Deep Learning." The project presents a hybrid deep learning framework for continuous cardiac monitoring using wearable photoplethysmography (PPG) signals. A one-dimensional U-Net is used to reconstruct ECG waveforms from PPG segments, while an attention-based CNN-LSTM model (Att-ECGNet) performs atrial fibrillation (AFib) detection directly on reconstructed ECG signals without requiring beat segmentation or spectrogram conversion. To protect sensitive physiological data, the classification pipeline incorporates Differential Privacy (DP) through per-sample gradient clipping and Gaussian noise injection, providing formal $(\varepsilon, \delta)$-DP guarantees during training. The framework is evaluated on multiple publicly available datasets, including the Cuffless Blood Pressure Estimation dataset, MIMIC-III waveform matched subset, and UMMC Simband dataset (Data uploaded here) . Experimental results demonstrate accurate ECG reconstruction, high AFib detection performance, and minimal performance degradation under differential privacy constraints. The repository includes datasets and codes used in the study.