Called the DiaBeats algorithm, new research suggests leveraging the machine learning-based algorithm, combined with ECG recordings, could help predict increased risk of developing prediabetes and type 2 diabetes.
Results of a new study provide evidence that suggests leveraging a combination of artificial intelligence and noninvasive imaging could provide a reliable, inexpensive screening method to improve detection rates of diabetes and prediabetes.
A study assessing the utility of electrocardiogram (ECG) recordings and a machine-learning algorithm, called the DiaBeats algorithm, results demonstrate use of the AI-integrated approach and, if externally validated, could be used to stratify individuals based on risk, which could prove useful in low-resource settings.
“In theory, our study provides a relatively inexpensive, non-invasive, and accurate alternative [to current diagnostic methods] which can be used as a gatekeeper to effectively detect diabetes and pre-diabetes early in its course,” wrote investigators. “Nevertheless, adoption of this algorithm into routine practice will need robust validation on external, independent datasets.”
As the notion of prediabetes as a benign condition has faded and the intrinsic risk associated with the disease has become recognized by the medical community, identification of low-cost, noninvasive methods for diagnosis or risk stratification for prediabetes and future diabetes has become paramount. With this in mind, a team from the Late Medical Research Foundation in Nagpur, India and M&H Research LLC sought to assess whether leveraging machine learning could identify prediabetes or diabetes from noninvasive cardiovascular imaging as the cardiovascular system bears some of the earliest signs of damage from the diabetic process.
To do so, investigators designed their study as an analysis using data from the Diabetes in Sindhi Families in Nagpur (DISFIN) study. A cross-sectional observational study aimed at estimating prevalence of type 2 diabetes in Sindhi families of Nagpur, 1262 of the 1462 individuals enrolled in the DISFIN study were included in the current analysis. From these 1262 individuals, investigators obtained data related to 10,461 time-aligned heartbeats recorded digitally.
This cohort had a mean age of 48 years and 61% were female. The overall prevalence of type 2 diabetes, prediabetes, and insulin resistance was 30%, 14%, and 35%, respectively. Investigators also pointed out hypertension, obesity, and dyslipidemia were present among 51%, 40%, and 36% of participants.
For the purpose of analysis, the 10,461 time-aligned heartbeats were split into a training set, a validation set, and an independent test set, which included 8892, 523, and 1046 beats, respectively. These recordings were processed with median filtering, band-pass filtering, and standard scaling. Investigators pointed out minority oversampling was performed to balance the training dataset prior to initiation of training and extreme gradient boosting was used to train the classifier that used the signal-processed ECG as input and predicted the membership to no diabetes, prediabetes, or type 2 diabetes classes.
Upon analysis, investigators found the DiaBeats algorithm predicted the classes with 97.1% precision, 96.2% recall, 96.8% accuracy, and 96.6% F1 score in the independent test set. Investigators noted a low calibration error (0.06) was observed with the calibrated model and the feature importance maps suggested that leads III, aVL, V4, V5, and V6 contributed most to the classification performance. Additionally, the predictions made by the algorithm matched the clinical expectations based on biological mechanisms of cardiac involvement in diabetes, according to investigators.
“Machine-learning-based DiaBeats algorithm using ECG signal data accurately predicted diabetes-related classes. This algorithm can help in early detection of diabetes and prediabetes after robust validation in external datasets,” investigators added.
This study, “Machine-learning algorithm to noninvasively detect diabetes and prediabetes from electrocardiogram,” was published in BMJ Innovations.