Model for Glycemic Level Detection using Heart Rate Variability in a Mexican Sample

Background: Type I and II diabetic patients must measure their glucose levels up to 10 times a day. The most common way is through an invasive blood test that causes pain, fear, discomfort and an important economic impact. Recent studies have shown a correlation between heart rate variability (HRV) and glucose levels, hypoglycemia, and hyperglycemia. This research aimed at designing a machine learning algorithm for glucose level detection using HRV features. Methods: We acquired physiological and clinical data from 5, yielding 100 observations. Then, we detected peaks from a photoplethysmography (PPG) signal and evaluated the peak detection accuracy. Afterwards, we measured multiple HRV-related features from such peaks. Then, data was curated, and relevant features were selected. After that, hyperparameters were optimized for 5 different estimators, and such values were used to create and evaluate predictive models. Results: The PPG peak detection algorithm performed efficiently, with an above 99% precision. The feature selection process yielded 20 features, 65% of them were HRV-related. The best model for glucose level detection was an ensemble between a Support Vector Machine, a Lasso and an Elasticnet model, with a mean absolute error of 16.24 mg/dL. Discussion: HRV features were very important to detect glucose levels. But fatigue, blood pressure and BMI were also included in the final model. We were limited by the size of the dataset, but our results are promising and HRV should be further explored as an indirect way of measuring blood glucose concentration. Additionally, and unlike similar studies in which HRV was measured from medical-grade electrocardiogram equipment, we measured HRV features from a PPG signal obtained using a commercial pulse oximeter. This would allow not only for a non-invasive glucose level measuring, but for it to be performed in a non-clinical setting.