A Machine Learning Approach for Diagnosis of Diabetes Through Fusion of PPG and ECG Signals

Currently, the worldwide prevalence of diabetes is about 537 million patients. Diabetes itself is not only fatal but also a cause of various diseases such as kidney and heart failures and other complications for instance neuropathy and ulcers. The prevailing methods to diagnose diabetes are all invasive (A1C, Fasting Glucose Plasma, etc.) and usually require from a few hours to a few days to provide accurate results. Hence, early diagnosis is essential to prevent the effects of diabetes and its related diseases. This work presents a non-invasive method to diagnose both type 1 and 2 diabetes. Here, we collected biosignals such as Electrocardiogram (ECG) and Photoplethysmography (PPG) using BIOPAC MP 36 and proposed a machine learning (ML) algorithm for detecting diabetes using a fusion of ECG & PPG signals. We utilized a band stop and notch filter for the removal of noise and motion artifacts contained within the acquired raw signals. Next, we extracted salient cepstral domain features from the filtered signals and employed a variety of classifiers for accurate diagnosis of diabetes. Fine-KNN classifier demonstrated an average accuracy of 98.46% and 99.03% for training and testing respectively by fusing information from ECG & PPG signals,