Data Quality Enhancement for Machine Learning on Wearable ECGs

With the exponentially growing amount of data collected by wearable medical devices, there is an expansion of opportunities to exploit machine learning methods in monitoring patient health and predicting events that may need medical attention. Wearable ECG devices provide a more comfortable alternative compared to the usual monitoring devices used in clinical settings, at the cost of inferior information density and signal quality. Notwithstanding, recent studies suggest that machine learning methods are able to work with wearable data, and for specific tasks, even medical grade devices are now available (e.g. detection of cardiac arrhythmias). This paper focuses on improving the quality of machine learning data, as part of an ongoing research of automating pipelines that allow for building robust models based on wearable ECG signals. While our efforts in general also consider the mainstream machine learning task of heartbeat classification, this work instead (exploiting the long-term ECG data) contributes to the development of models to learn features much longer than a single heartbeat. For the extraction of such features, R-peak detectors are considered and evaluated with respect to sensitivity and robustness to noise. Noise stress tests show that a combination of global filters can improve the performance of detectors, efficiently dealing with typical noises almost always present in wearable ECG signals. Also, local noise detection models are demonstrated to be promising methods to handle heavy noises that can not be resolved by the global filters.