Predicting High Levels of Blood Glucose through Heart Rate Variability and Machine Learning

As the current literature shows, there is a correlation of Heart Rate Variability (HRV) and the levels of glucose in blood (BGL), which can be used when assessing diabetes related risks. Commonly, a blood glucose meter is used as an invasive, and often expensive, tool to measure BGL at multiple times of day; however, there are studies showing noninvasive solutions that use the principles of HRV instead. Measuring BGL with a pulse oximeter and HRV is a beneficial, cost-effective, non-invasive method that represents an easier, less-frightening approach. We developed a methodology based on information obtained from our previous works, to create predictive models based on machine learning methods such as Twin Support Vector Machines (TWSVM) and boosted Bayesian Additive regression Trees (xBART). Utilizing a total of 58 variables generated by HRV variables and pulse oximeter information, we computed cross-validation tests, and a general train-test split for both models. The performance of the models was computed with balanced accuracy and weighted F1-score metrics, resulting in promising predictions that ranged upwards of 72.2% to 82.1% for the accuracy. As for the weighted F1-score, our results ranged from 75.0% to 76.5%.