Deep learning model optimization for the classification of 27 cardiac arrhythmias: a study on performance improvement and clinical feedback integration

Cardiovascular disease poses a significant health risk and remains one of the leading causes of death worldwide. The 12-lead electrocardiogram is a comprehensive and widely accessible diagnostic tool for identifying cardiac abnormalities. Early and accurate diagnosis allows for timely treatment and intervention, helping to prevent severe complications. It is therefore crucial to have automated classification tools capable of recognizing these cardiac alterations, streamlining clinical practice, which is currently very time-consuming. This thesis contributes to this context by building on a solution developed by a team ranked third in the PhysioNet/Computing in Cardiology Challenge 2020, which aimed to classify 27 cardiac arrhythmias using a new scoring metric. This metric awards partial credit for misdiagnoses that result in similar outcomes or treatments as the correct diagnosis, according to cardiologists. Starting from the original model, a squeeze-and-excite ResNet ensemble using the same signal from eight truncated leads of varying lengths, several preprocessing steps were applied to improve performance. Furthermore, clinical expertise was incorporated through collaboration with a cardiologist from a hospital in Turin. One of the most promising modifications, integrated into the final solution, was the addition of features that describe the patient’s condition, such as gender and age, indicating that more patient-specific information leads to better outcomes and more tailored treatment approaches. After multiple refinements, the final model consists of two phases: the first is a binary model that discriminates between healthy and altered classes, achieving an accuracy of 88% and a macro F1 score of 93%. The second phase distinguishes 19 classes by grouping certain arrhythmias and integrating clinical knowledge. The results are promising, with overall performance improvements on both the validation set and hidden test set. The integration of clinical data has proven to be crucial, highlighting the importance of close collaboration with field experts to refine the model and ensure its clinical applicability in real-world settings.