Abstract
In this pilot study, we evaluate the diagnostic accuracy and feasibility of utilizing smartphone-connected electrophysiological biosensors for measuring blood glucose level (BGL) in a type-2 diabetic (T2D) patient using machine learning (ML). Patients' electrophysiological parameters are measured using commercially available, smartphone operated/embedded electrocardiography (ECG) and photoplethysmography (PPG) biosensors; simultaneously, we measure the BGL of the patient using a minimal invasive (MI) biosensor mounted on the patient's arm. Statistical correlation analysis confirms that a modest correlation exists between the extracted ECG and PPG features and the measured BGL. The efficacy of ML models trained using the regression tree algorithm utilizing ECG and PPG extracted features for predicting the BGL is investigated. Experimental results show that ECG and PPG feature trained ML models can successfully predict BGL upto ± (8-33 mg/dL) during fasting, pre-prandial, and post-prandial intervals.