Previously published photoplethysmography-(PPG) based non-invasive blood glucose (NIBG) measurements have not yet been validated over 500 subjects. As illustrated in this work, we increased the number subjects recruited to 2538 and found that the prediction accuracy (the ratio in zone A of Clarke’s error grid) reduced to undesirable 60.6%. We suspect the low prediction accuracy induced by larger sample size might arise from the physiological diversity of subjects, and one possibility is that the diversity might originate from medication. Therefore, we split the subjects into two cohorts for deep learning: with and without medication (1682 and 856 recruited subjects, respectively). In comparison, the cohort training for subjects without any medication had approximately 30% higher prediction accuracy over the cohort training for those with medication. Furthermore, by adding quarterly (every 3 months) measured glycohemoglobin (HbA1c), we were able to significantly boost the prediction accuracy by approximately 10%. For subjects without medication, the best performing model with quarterly measured HbA1c achieved 94.3% prediction accuracy, RMSE of 12.4 mg/dL, MAE of 8.9 mg/dL, and MAPE of 0.08, which demonstrates a very promising solution for NIBG prediction via deep learning. Regarding subjects with medication, a personalized model could be a viable means of further investigation.
A Deep Learning Approach to Diabetic Blood Glucose Prediction