Academics have applied for a patent describing how a neural network can detect low blood-sugar levels by analyzing heartbeat patterns rather than a blood sample.
Keeping track of glucose levels is annoying and painful. Multiple times per day, diabetics have to prick their finger, place the small drop of blood on a test strip, and insert that strip into a glucometer to get a reading, and then dial up their insulin dosage, or eat or drink carbohydrates, as necessary. In the US, at least, these strips aren't cheap, thanks to the healthcare system.
The AI-based method developed by the team, however, is non-invasive. It only requires people to wear a device that can measure electrocardiograms (ECG), recordings of heartbeats made by sensors placed on the skin. Abnormal blood glucose levels can affect ECG readings; high levels of sugar lead to rapid heart rates, whereas low levels correlate to low heart rates.
The ECGs are then processed by a convolutional neural network (CNN) and a recurrent neural network (RNN) to flag up episodes of nocturnal hypoglycemia, a condition where glucose levels below a normal range during sleep.
Described as a “pilot study,” the researchers recruited four volunteers to wear devices that measure both ECGs and a non-invasive continuous glucose monitor (CGM). Over the course of up to 14 days, they studied each person’s pulse at times when their heart rates were normal and when they were affected by nocturnal hypoglycemic events. The data from the ECG and CGM were correlated and used to train the CNN and RNN to predict when blood glucose levels dip below normal levels from an individual’s heart rate.
Some of the ECG readings were held back for testing, and the results showed the team’s neural networks were on average accurate roughly 82 per cent time.
“Our approach enables personalized tuning of detection algorithms and emphasizes how hypoglycaemic events affect ECG in individuals,” said Leandro Pecchia, co-author of the paper and an associate professor of biomedical engineering at the University of Warwick, England. “Based on this information, clinicians can adapt the therapy to each individual.”
Why build your own cancer-sniffing neural network when this 1.3 exaflop supercomputer can do if for you?READ MORE
But before any diabetics out there get their hopes up over such a device, the team admitted their patent has to go through much more clinical testing. Firstly, not only is their research sample size small, but none of the participants had type 1 or type 2 diabetes.
“Our study concerned the detection of nocturnal non-induced low glucose levels in healthy individuals; several clinical studies showed that cardiac changes could have different intensities in healthy, type 1 and type 2 diabetic persons,” the paper said.
So far, the results do show that applying deep learning on ECG can detect low blood glucose events and that training on personalized data makes it more effective for individuals. The goal is to eventually develop a device for diabetics that alerts them whenever their glucose levels dip to dangerous levels in their sleep.