PPG Signal Processing and Artificial Intelligence based Non-Invasive Diabetes Identification

The global prevalence of diabetes, affecting more than 500 million individuals. It underscores the urgent need for frequent blood glucose monitoring. Traditional invasive methods cause discomfort and pose challenges such as slow wound healing for diabetic patients. Therefore, precise non-invasive diabetes identification is becoming essential. The aim of this paper is to develop an intelligent Photoplethysmography (PPG) signal based non-invasive automated identification of diabetes. The considered raw PPG signals are pre-processed. The pre-processing is composed of the smoothening window application to avoid any spectral leakage during segmentation. Then digital band-pass filtering is applied for noise removal and base line-restoration. Afterward, the Z-score normalization is used to diminish the impact of multi-subject data. The class imbalance problem is tackled using a uniform distribution random up-sampling approach. Afterward, the pre-processed PPG signals are analyzed and decomposed in sub-bands using the wavelet decomposition (WD). The pertinent features are extracted from sub-bands by performing the statistical analysis of sub-band coefficients. The importance of features is verified performing the analysis of variance (ANOVA) test and selecting the features with p-value<0.05. The automated categorization is carried out using robust machine learning (ML) and ensemble learning (EL) classifiers. The applicability is tested for the case of a real multi-class and multi-subjects PPG dataset. The devised solution achieves the highest average classification accuracy of 96.60%.