2 June 2026 in IEEE Sensors Journal
Photoplethysmography (PPG) captures valuable information related to vascular and hemodynamic changes due to various cardiovascular and hemodynamic disorders. Recent research highlights PPG as a promising modality for non-invasive type-II diabetes detection. However, published models in this area are either computationally heavy or lacks interpretability. Secondly, the trained model parameters being static, performance may degrade over time due to varied physiological conditions. This study proposes a new method of optimal frequency along with ensemble empirical mode decomposition (EEMD) derived intrinsic mode function (IMF) selection based on a new parameter named discriminative index (DI). Using publicly available PPGBP dataset, an accuracy of 91% for hypertensive subjects and 92% for pre-hypertensive and non-hypertensive subjects were obtained using random forest classifier. The algorithm also assesses the reliability of the prediction using a KL-divergence and provide a decision whether a retraining is required for selection of the optimal feature for new test subjects. The end-to-end pipeline was also deployed in a resource constrained hardware ARM v8 controller, with a low memory requirement of 2.73 MB and detection latency of 29.4 s to process 2.1 s PPG snippet. The proposed method can be useful for non-invasive diabetes sensor development towards improved diabetes management.