Machine-learning algorithm detects diabetes class from ECG

A machine-learning algorithm dubbed DiaBeats demonstrates accuracy in identifying diabetes and prediabetes noninvasively using electrocardiogram (ECG) signal data, reports a study.

“In theory, our study provides a relatively inexpensive, noninvasive and accurate alternative which can be used as a gatekeeper to effectively detect diabetes and prediabetes early in its course,” the researchers said. “Nevertheless, adoption of this algorithm into routine practice will need robust validation on external, independent datasets.”

This study used data from the Diabetes in Sindhi Families in Nagpur study of ethnically endogenous Sindhi population from central India. Clinical data from 1,262 individuals and 10,461 time-aligned heartbeats recorded digitally were included in the final dataset, which was divided into a training set, a validation set, and an independent test set (8,892,523, and 1,046 beats, respectively).

The researchers processed ECG recordings with median filtering, band-pass filtering, and standard scaling. They also carried out minority oversampling to balance the training dataset prior to training initiation. Finally, extreme gradient boosting (XGBoost) was applied to train the classifier who used the ECG data as input and predicted diabetes classes based on the American Diabetes Association criteria.

Training was smooth and fast, with convergence achieved within 40 epochs. Diabetes had a nearly 30-percent prevalence, while that for prediabetes was about 14 percent. The DiaBeats algorithm predicted the diabetes classes with 97.1-percent precision, 96.2-percent recall, 96.8-percent accuracy, and 96.6-percent F1 score in the independent test set. [BMJ Innov 2023;9:32-42]

Calibration error in the calibrated model was low (0.06). In the feature importance maps, the elements that contributed most to the classification performance were leads III, augmented Vector Left, V4, V5, and V6. Notably, the predictions made by the algorithm were consistent with the clinical expectations based on the biological mechanisms of cardiac involvements in diabetes.

Detection by ECG

Using ECG to detect diabetes and prediabetes has a strong biological basis, and the need to assess cardiac status in patients with type 2 diabetes has been long acknowledged, according to the researchers. [Diabetologia 2008;51:1581-1593]

For instance, stimulation of the connective tissue growth factors, fibrosis, accumulation of advanced glycation end-products, and overall stiffness of the heart muscle—events that indicate diabetic cardiomyopathy—manifest in the basal and septal areas of the left ventricle. [Endocr Rev 2004;25:543-567; Am J Physiol Heart Circ Physiol 2009;297:H2109-2119; Circulation 2009;120:1633-1636]

“Interestingly, these changes in heart structure and function begin very early in disease and are known to occur with sustained hyperglycaemia in the prediabetic range as well,” the researchers said. [Int J Mol Sci 2019;20:3299]

“Therefore, ECG can be of value early during diabetes/prediabetes to detect the subtle, subclinical, concomitant, and characteristic cardiac involvement,” they added.

XGBoost

Compared with other machine and deep learning models, the XGBoost-based DiaBeats classifier was deemed the “best performing algorithm,” according to the researchers, noting its ensemble approach based on decision-tree analytic framework.

Several advantages of XGBoost algorithm included hardware optimization and speed, a parallelized tree-building process, tree pruning using a depth-first approach, and L1 and L2 regularization to reduce or prevent overfitting.

Two recent reviews demonstrated the value of this algorithm in clinical classification tasks. [Int J Med Inform 2022;159:104679; Emerg Top Life Sci 2021;doi:10.1042/ETLS20210246]

“Early detection is of crucial importance for prevention of type 2 diabetes and prediabetes,” the researchers said. “Diagnosis of these conditions relies on the oral glucose tolerance test and haemoglobin A1c estimation, which are invasive and challenging for large-scale screening.” [Diabetes Metab Res Rev 2014;30:654-658]