Browse By

A noninvasive method of detecting low-glucose events via ECG with AI

Written by JuWon Park

Hypoglycemia, a condition caused by low blood glucose levels, is usually an indicator of other health problems. The most common cause of this condition is as a side effect of diabetes treatment (“Hypoglycemia,” n.d.). Blood sugar levels can be tested with a blood glucose meter, which requires the patient to prick his/her finger. 

Alternative devices such as Continuous Glucose Monitoring Devices (CGMs) measure blood glucose levels in real-time, but these devices are still invasive and have limited reliability. CGMs work by measuring blood glucose from interstitial fluid using a small needle and transmitting the information to a pager-like monitor every so often. However, CGMs need to be calibrated with an invasive finger-prick blood glucose test multiple times a day (“How Does a Continuous Glucose Monitor Work?”, 2019).

Previously, scientists attempted to use an electrocardiogram (ECG)- based hypoglycemic detection methods, but the differences between individuals (intersubjective differences) made it difficult. To clarify, an ECG is a test that measures the electrical activity of the heart. When the heart beats, it sends an electrical impulse through the heart. The ECG will show the timing of which the impulse travels from the top to lower chambers of the heart (“Electrocardiogram (ECG or EKG)”, n.d.). Attempts like these that have failed due to inaccurate results have attracted attention towards personalized medicine that can make treatments and diagnostics more accurate based on individualized data such as their lifestyle, physiology, and genome. 

Dr. Leandro Pecchia’s team at the University of Warwick was able to use AI to develop an off-the-shelf non-invasive sensor that can detect hypoglycemia from ECG signals (“AI can detect low-glucose levels,” 2020). Two pilot studies showed that its performance was similar to that of current CGM performance. 

The Warwick AI model is able to detect hypoglycemic events because of its personalized, novel approach that takes into account different cardiac cycles. Previous studies failed to detect hypoglycaemic events because they used cohort (or group) ECG-data that did not account for inter-subject differences (“AI can detect low-glucose levels,” 2020). The study, instead, trains the model with ECG data from an individual during the first days and tests the model using data from the same individual (Porumb, Stranges, Pescape, & Pecchia, 2020).

For each participant, the recording nights were split into 2 separate datasets for training and testing the model, ensuring that every dataset contained nights with low blood glucose events (Porumb, Stranges, Pescape, & Pecchia, 2020). 

Specifically, continuous ECG and glucose readings from healthy volunteers from midnight to 9 AM were measured. Data was collected at night because of ECG circadian changes and a lack of low glucose events during the day. Lower glucose level events were when the glucose concentration was less than 4 mmol/L. They also processed the heartbeats to filter out noisy, or inaccurate ECG readings. In their final method, the model is able to automatically detect low glucose events in an individual’s ECG changes (Porumb, Stranges, Pescape, & Pecchia, 2020).

The accuracy of the model provides strong evidence that nocturnal ECG readings can be used to build real-time low glucose event monitoring systems. The proposed method has the potential to create long-term improvements in the clinical outcomes of diabetic patients and many others.

Works Cited

Dansinger, M. (2019, December 1). How does a continuous glucose monitor work? Retrieved from

Electrocardiogram (ECG or EKG). (n.d.). Retrieved from 

University of Warwick. (2020, January 13). AI can detect low-glucose levels via ECG without fingerprick test. ScienceDaily. Retrieved January 26, 2020 from

Porumb, M., Stranges, S., Pescapè, A., & Pecchia, L. (2020). Precision Medicine and Artificial Intelligence: A Pilot Study on Deep Learning for Hypoglycemic Events Detection based on ECG. Scientific Reports, 10(1), 1-16.


Leave a Reply

Your email address will not be published. Required fields are marked *