ECG reading by artificial intelligence can identify cardiovascular autonomic neuropathy in diabetes

Aims: To assess the ability of machine learning algorithms to detect cardiac autonomic neuropathy (CAN) via analysis of 12-lead, 10-s electrocardiogram (ECG) tracings.

Methods: This observational, single-centre study involved extraction of motifs (i.e., frequently occurring patterns) and discords (i.e., rare/anomalous patterns) in addition to detection of long-term dependencies within ECG recordings through long-short term memory (LSTM) networks. Extracted features were used to train a classification model named support vector machine (SVM), whose performance was tested by a 10-fold cross validation with one tenth of the dataset serving as the test set and the remaining nine tenths serving as the training and validation sets.

Results: Among 205 patients with type 1 and type 2 diabetes (mean age 57 ± 17 years, 54% females) included in the study, 100 were diagnosed with CAN by means of cardiovascular autonomic reflex tests (CARTs), including 38 patients with definite/severe CAN (dsCAN) and 62 patients with early CAN (eCAN). The model combining motifs and discords exhibited the best performance and was able to detect dsCAN with an accuracy of 0.92, an F1 score of 0.92, a precision of 0.91, and an area under the receiver-operating characteristic curve (AUC) of 0.93 (95% CI 0.91-0.94). On the other hand, it was less reliable in identifying any stage of CAN (0.65 accuracy, 0.68 F1 score, 0.75 recall, 0.68 precision, 0.68 AUC).

Conclusions: a machine learning approach, particularly identifying motifs and discords within ECGs, accurately detected dsCAN in patients with diabetes.

Comments: Cardiac autonomic neuropathy (CAN) frequently remains undiagnosed among patients with diabetes despite its significant morbidity and mortality. One reason for this is the fact that CARTs, which are the gold standard diagnostic procedures, are cumbersome, time-consuming and not available in all clinical settings. It is therefore crucial to devise more convenient screening tools aimed at extending diagnosis of CAN and allowing for earlier intervention. This paper provides promising results as to the use of machine learning techniques capable of carrying out ultra-short-term heart rate variability measurements to detect CAN. However, the accuracy of these tools in identifying early CAN is limited compared to definite/severe CAN. As patients with early CAN are less likely to be referred to CARTs due to their less severe clinical manifestations, they would benefit the most from effective screening methods.

The real-world implementation of such machine learning algorithms requires validation and standardization across diverse populations and clinical settings. To date, it is clear that artificial intelligence-enhanced ECG reading cannot replace CARTs but might represent a tool for large-scale screening. Further studies are needed to assess whether overall screening accuracy can be improved by combining ECG analysis with clinical risk scoring systems in order to identify CAN at all stages.

Pietro Pertile

Reference. Irlik K, Aldosari H, Hendel M, Kwiendacz H, Piaśnik J, Kulpa J, Ignacy P, Boczek S, Herba M, Kegler K, Coenen F, Gumprecht J, Zheng Y, Lip GYH, Alam U, Nabrdalik K. Artificial intelligence-enhanced electrocardiogram analysis for identifying cardiac autonomic neuropathy in patients with diabetes. Diabetes Obes Metab. 2024 Jul;26(7):2624-2633. doi: 10.1111/dom.15578. Epub 2024 Apr 11. PMID: 38603589.

https://dom-pubs.pericles-prod.literatumonline.com/doi/10.1111/dom.15578