Abstract
Atrial fibrillation (AFib) often escapes detection in short-duration electrocardiograms (ECGs) and conventional Holter monitors. This project develops a low-cost, consumer-grade wearable screening system that pairs with BITalino HeartBIT single-lead ECG sensor with an Android smartphone for real-time, on-device AFib detection. The system streams at 1 kHz, applies lightweight band-pass and notch filtering, detects R-peaks, and derives RR-interval (RR) based Heart-Rate Variability (HRV) features directly on the device. These QRS complexes, RRs and features are processed using a principal component analysis (PCA) reduction and evaluated with multiple machine-learning models, including a RBF-kernel support vector machine (SVM), a 1-D convolutional neural network (CNN), and a long short-term memory (LSTM) network. Based on literature and empirical testing, the final model uses a compact PCA + SVM pipeline exported to ONNX for mobile inference efficiency. Evaluations across combines PhysioNet (AFDB, LTAFDB, NSRDB) and pilot BITalino recordings (383 labeled RR windows) yields approximately 93.5% accuracy, 97.6% sensitivity, and 88.6% specificity on a subject-wise held-out test set. The Android implementation incorporates signal-quality checks, a sliding-window consensus logic for accurate decisions, and a user interface, which displays a live ECG preview, recording, and AFib alerts. This is a demonstration that transparent hardware and lightweight ML models can provide affordable, privacy-preserving, real-time AFib screening on everyday smartphones, while highlighting that the long-term validation, electrode reliability, and extended field deployment are still challenges.