Currently, Continuous Glucose Monitors (CGM) measure glucose in interstitial fluid using an invasive sensor with a little needle, which sends alarms and data to a display device. In many cases, they require calibration twice a day with invasive finger-prick blood glucose level tests.
Dr. Leandro Pecchia’s team at the University of Warwick has published results in a paper titled ‘Precision Medicine and Artificial Intelligence: A Pilot Study on Deep Learning for Hypoglycemic Events Detection based on ECG’ in the Nature Springer journal Scientific Reports proving that using the latest findings of AI (i.e., deep learning), they can detect hypoglycemic events from raw ECG signals acquired with off-the-shelf non-invasive wearable sensors.
Two pilot studies with healthy volunteers found an average sensitivity and specificity of approximately 82% for hypoglycemia detection, which is comparable with the current CGM performance.
Dr Leandro Pecchia from the School of Engineering at the University of Warwick commented, “Taking fingerpick during the night certainly is unpleasant, especially for patients in pediatric age. Our innovation consisted of using artificial intelligence to automatically detect hypoglycemia via ECG beats. This is relevant because ECG can be detected under any circumstances, including while the user is sleeping.”
The Warwick model highlights how the ECG changes in each subject during a hypoglycemic event. The figure below is an exemplar. The solid lines represent the average heartbeats for two different subjects when the glucose level is normal (green line) or low (red line). The red and green shadows represent the standard deviation of the heartbeats around the mean. A comparison highlights that these two subjects have different ECG waveform changes during hypo events. In particular, Subject 1 presents a visibly longer QT interval during hypo, while the subject 2 does not. The vertical bars represent the relative importance of each ECG wave in determining if a heartbeat is classified as hypo or normal.
From these bars, a trained clinician sees that for Subject 1, the T-wave displacement influences classification, reflecting that when the subject is in hypo, the repolarization of the ventricles is slower.
In Subject 2, the most important components of the ECG are the P-wave and the rising of the T-wave, suggesting that when this subject is in hypo, the depolarization of the atria and the threshold for ventricular activation are particularly affected. This could influence subsequent clinical interventions.
This result is possible because the Warwick AI model is trained with each subject’s own data. Intersubjective differences are so significant, that training the system using cohort data would not give the same results. Likewise, personalized therapy based on this system could be more effective than current approaches.
“The differences highlighted above could explain why previous studies using ECG to detect hypoglycemic events failed. The performance of AI algorithms trained over cohort ECG-data would be hindered by these inter-subject differences,” added Pecchia. “Our approach enables personalized tuning of detection algorithms and emphasizes how hypoglycemic events affect ECG in individuals. Based on this information, clinicians can adapt the therapy to each individual. Clearly more clinical research is required to confirm these results in wider populations. This is why we are looking for partners.”