Machine Learning-Based Multivariate Classification of Cardiac Autonomic Neuropathy Via Imbalanced ECG Recordings

Cardiac Autonomic Neuropathy (CAN) is a frequent complication of diabetes that presents a diagnostic challenge in clinical settings. Ewing tests, comprising five assessments of blood pressure and heart rate, have traditionally been used to categorize the severity of CAN. This paper investigates the application of a machine learning (ML) approach that leverages patient demographic information and electrocardiogram (ECG) recordings to automatically classify CAN dysfunction without any of the Ewing tests. Additionally, the ML model being proposed addresses the challenges posed by imbalanced datasets both at the data level, through the utilization of synthetic minority over-sampling technique (SMOTE), and at the algorithmic level, by employing the focal loss function. The effectiveness of the proposed techniques is reflected in the ROC curve achieving an area under curve (AUC) value of 92%. Moreover, the model is able to produce 80% accurate prediction of the development of CAN. However, for enhancing system's classification performance, two Ewing tests are supported as additional features. A higher performance is achieved as indicated by accuracy and F1-score values of 88% and 88%, respectively.