Online anomaly detection for long-term ECG monitoring using wearable devices

Many successful algorithms for analyzing ECG signals leverage data-driven models that are learned for each specific user. Unfortunately, a few algorithmic challenges are still to be addressed before employing these models in wearable devices, thus enabling online and long-term monitoring. In particular, since the heartbeats morphology changes with the heart rate, models learned in resting conditions need to be adapted to analyze ECG signals recorded during everyday activities. We propose an online ECG monitoring solution where normal heartbeats of each specific user are modeled by dictionaries yielding sparse representations, and heartbeats that do not conform to this model are detected as anomalous. We track heart rate variations by adapting the user-specific dictionary with a set of user-independent, linear, transformations. Our experiments demonstrate that these transformations can be successfully learned from a public dataset of ECG signals and that, thanks to an optimized anomaly-detection algorithm, our solution enables online and long-term ECG monitoring.