IDENTIFICATION OF SIGNIFICANT TEMPORAL FEATURES FROM PHOTO PULSE PLETHYSMOGRAM FOR PREDICTION OF HYPERGLYCEMIA

The increase in diabetic class population is witnessed every year and hence frequent need of checking blood glucose levels is highly required in both diabetic and healthy subjects. This requirement has led many researchers to develop efficient, accurate and easy estimation of blood glucose level without painful blood pricking or extraction. In line with this objective, this study is an effort for waveform analysis of autonomic response of photo pulse plethysmogram (PPG) and to evaluate PPG features that may provide better non-invasive detection of the disease. A total of 300 (150 diabetic and 150 control) subjects aged more than 18 years are considered for this study. Relevant patient details were obtained, and finger PPG recording in digital form was performed for 5 min with the digitization sampling rate of 200 sample points per second. After signal baseline correction, features are calculated from averaged normalized PPG and its first and second derivatives. A total of 41 features were derived from normalized PPG pulse and its derivatives for the assessment of hyperglycemia. But in practice, computation of many features is time-consuming and truly a tedious job for any model design. Further, model performance and accuracy are affected significantly. Applying six different statistical and machine learning methods for the identification of important features, 20 features were identified out of 41 features whose values are highly significantly different for control and study (diabetic) groups. This work was focused on feature extraction stage for a PPG-based diagnostic system. Though our study is aimed to identify significant features that can discriminate between diabetic and control groups non-invasively, these features have the potential for much clinical assessment and other pathological disorders without any doubt.