Non-invasive Diabetes Mellitus Detection System using Machine Learning Techniques

This work presents an automated diabetes mellitus detection system (DMDS), based on wrist photoplethysmography (PPG) signal and physiological parameters. The PPG signal with an average duration of 2.5 minutes is obtained using the handle Empatica E4 Wristband from 217 patients. The mel frequency cepstral coefficients (MFCC) features are extracted from 5 second segments of the PPG signal. The extracted features and physiological parameters constitute the input for machine learning (ML) systems. K-nearest neighbors (KNN) and Support Vector machine (SVM) algorithms are used for classification. 83.87% and 84.49% classification accuracy is achieved with KNN and radial basis function (RBF) Kernel SVM based DMDS respectively. Further principal component analysis is used on the input feature set to the SVM classifier which provides 7.79% improvement in the performance. The performance of the developed systems is also analysed using entropy triangle. Results reveal the effectiveness of proposed DMDS for non-invasive DM detection. The designed wrist band identifies the diabetic and pre-diabetic cases in real time on the basis of short duration PPG signal and physiological parameters.