Towards Finger Pulse Photoplethysmogram Based Non-invasive Classification of Diabetic versus Normal

There has been a markedly increase in the number of published articles and citations describing both basic and applied research on PPG (Photo-Plethysmogram) over the past 50 years for being low cost, easy, safe, and non-invasive tool. This article is the effort for classification study for diabetic versus normal subjects based on 15 time domain features of PPG wave form using the machine-learning Random Forest Classifier algorithm. Also, first and second derivative features of PPG pulses are calculated and identified for more accuracy. The presented study is based on the evaluation through 5 min fingertip PPG recordings for 124 subjects diagnosed as diabetic and 166 subjects categorized as non-diabetic group. We achieved an overall accuracy as 97.7% for two classification groups DM (diabetic mellitus) and non-DM. For non-diabetic group Precision, Recall and F1 score are observed as 0.95, 1.00, and 0.98, respectively. Similarly for diabetic group the values are Precision (1.00), Recall (0.96) and F1score (0.98), respectively. The results have been cross validated with optimized classifier parameters which ensure the robustness of the model. Since the presented study is based on real-time recorded PPG waveforms and some of the features used for this classification job are new and significant, which makes this study an important work.