Type2 Diabetes Classification from Short Photoplethysmogram Signal Using Multiple Domain Features and Machine Learning Techniques

Purpose Type 2 diabetes is a rapidly growing chronic disease that needs early detection. If untreated in the initial stage, diabetes-related problems such as retinopathy, neuropathy, and cardiac complications arise in the body. It must be detected early in order to receive proper treatment, which can prevent the condition from progressing to severe complications. This study proposed a non-invasive type 2 diabetes classification system using short PPG signals through various machine learning (ML) techniques. Method Our dataset includes 80 healthy and 80 type2 diabetic patients. From each subject, we extracted total 118 features from four different domains (time, frequency, demographic-Time domain, and demographic domain). We applied various ML techniques on 68 statistically significant features with or without feature selection technique (FST). Results Without using the FST, we obtain the highest accuracy of 99.88% and 99.89% using Fine K-nearest neighbor (KNN) and Subspace Ensemble of K-nearest neighbor (SE-KNN) techniques respectively. With FST we achieved an accuracy of 97.083% by Cubic support vector machine (C-SVM), with only 9 selected features. Conclusion Our successful implementation of the ML classifier can be implemented in wearable or smart applications which will enable people to assess early type2 diabetes in a more convenient way.