An artificial intelligence model using standard electrocardiograms showed stronger discrimination at a high threshold for patient-reported peripheral neuropathy symptoms than at a broader threshold in adults with diabetes, according to a single-center study published in Diabetes, Obesity and Metabolism. 

The researchers retrospectively analyzed prospectively collected data from 640 adults with diabetes who attended an outpatient clinic in Zabrze, Poland, between July 2022 and April 2025. Each patient completed the Michigan Neuropathy Screening Instrument questionnaire (MNSI-Q) and underwent a standard 12-lead electrocardiogram (ECG). Questionnaire scores defined the study outcome. A structured neurological examination was available but was not included in the primary definition; nerve conduction testing was not used as the reference standard. 

The analysis used 10 seconds of ECG data per patient. The researchers extracted two types of waveform features—recurring patterns (motifs) and anomalous patterns (discords)—then tested three machine learning classifiers: XGBoost, a support vector machine, and ridge regression. They evaluated the models using repeated nested cross-validation within the study cohort but did not validate them in an external cohort or test them in a live clinical workflow. 

The primary analysis classified 95 patients (15%) as meeting the higher symptom threshold, using an MNSI-Q score of at least seven. A secondary analysis used a score of at least four, a broader threshold met by 281 patients (44%).  

For the higher symptom threshold, the XGBoost model using both motifs and discords had the highest overall performance, achieving an area under the receiver operating characteristic curve (AUROC) of 0.89, 89% accuracy, 93% sensitivity, and 85% positive predictive value. In comparison, the support vector machine trained on both features achieved an AUROC of 0.88, 87% accuracy, 83% sensitivity, and 93% positive predictive value. Ridge regression using both feature types had an AUROC of 0.63, 62% accuracy, 67% sensitivity, and 62% positive predictive value. The study did not report specificity or negative predictive value for these models. 

At the broader threshold, XGBoost using both motifs and discords had an AUROC of 0.64, 64% accuracy, and 69% sensitivity. The support vector machine using both feature types also had an AUROC of 0.64 and 64% accuracy. Its sensitivity was higher than XGBoost’s (80% vs 69%), but its positive predictive value was lower (60% vs 64%). Ridge regression again performed the lowest with an AUROC of 0.59, 59% accuracy, and 62% sensitivity. These metrics describe classification against questionnaire scores, not examination- or nerve conduction–confirmed neuropathy. 

Patients above the higher questionnaire threshold had a longer median diabetes duration than those below it (14 vs 10 years) and higher glycated hemoglobin levels. They also had a higher prevalence of hypertension, coronary artery disease, and diabetic retinopathy. 

In a separate analysis involving 561 patients with complete covariate data, adding the ECG-derived score to a model containing age, sex, cardiovascular conditions, kidney function, glycated hemoglobin, and diabetes duration increased the AUROC from 0.706 to 0.904. This comparison was conducted within the study cohort and does not establish clinical benefit. 

The authors cited several limitations, including the single-center design, symptom-based case definition, imbalance between patients above and below the higher threshold, and lack of comparison with other ECG-based approaches. They said unmeasured factors could still confound the association between the ECG score and symptoms. Their decision-curve analysis also could not reliably quantify clinical utility.  

For future research, the authors called for prospective external validation across different patient populations and ECG systems, objective neuropathy assessment in at least a subset of patients, and a decision-curve analysis in the external validation study. 

“With further validation, this method could serve as an accessible, supplementary screening aid to help identify high-risk patients during routine cardiovascular assessment,” the authors wrote.  

Hanna Kwiendacz, MD, PhD, Uazman Alam, PhD, Janusz Gumprecht, MD, Gregory Y. H. Lip, MD, and Katarzyna Nabrdalik, MD, reported industry payments, professional roles, or research support. The remaining authors reported no conflicts of interest.