Medical Signal Analysis For Early ECG Classification Using Machine Learning Models

Cardiac conditions are among the primary factors contributing to worldwide mortality. Therefore, early detection and warning for patients play a vital role that can help them avoid potential cardiovascular risks in a timely manner. In this article, we introduce two approaches for early detection of cardiac arrhythmias and conduct various experiments on a wide range of different machine learning models ranging from traditional statistical to deep learning models. This provide a comprehensive picture of the overall performances of these models on the early cardiac arrhythmia prediction task for future research in other areas. In the first approach, features are extracted from historical electrocardiogram data to serve as inputs for the models, while the second approach involves early classification from future heartbeats simulated through regression on the historical electrocardiogram. For each approach, we propose additional methods expected to enhance sensitivity in anomaly prediction. The results achieved are as follows: the first approach, using the DuoHF-CNN network that we propose, improved abnormality prediction sensitivity by up to 55%, while maintaining a harmonious normal class performance at 85%. Regarding the second approach, the transfer learning models that we experimented with revealed inconsistent performance. While some models demonstrated reasonably good results, achieving a normal class sensitivity of 63%, the abnormality detection rate was at 52%.