Cardiac comorbidity role in the diagnosis of diabetes mellitus using single lead electrocardiogram utilizing machine learning model

Abstract

BACKGROUND

Cardiovascular disease alters electrocardiographic morphology and may obscure diabetes-related electrical signatures, yet its effect on electrocardiogram (ECG)-based detection of type 1 and type 2 diabetes is poorly defined.

AIM

To investigate whether cardiac comorbidity differentially influences machine-learning detection of the two diabetes types from single-lead ECG.

METHODS

In this study, 2591 observations remained after exclusion of an erroneous age entry and rerunning the analytical pipeline. Twenty-six XGBoost models were evaluated separately for type 1 and type 2 diabetes within subgroups defined by seven cardiac-status criteria. Primary models used 47 single-lead ECG features and five anthropometric variables and were evaluated with five-fold stratified cross-validation. A post hoc seven-predictor sensitivity analysis assessed model complexity. Area under the receiver operating characteristics curve (AUC) uncertainty was estimated using 2000 bootstrap resamples of out-of-fold predictions, and ΔAUC comparisons were adjusted using the Benjamini-Hochberg procedure. Calibration was assessed using Brier score, calibration intercept and slope, reliability plots, and decision-curve analysis. Validation was observation-level because a persistent patient identifier was not available in the analytical export.

RESULTS

After correction and full rerun, median AUC was 0.693 for type 1 diabetes (range: 0.660-0.804) and 0.661 for type 2 diabetes (range: 0.543-0.716). No cardiac-pathology ΔAUC comparison remained statistically significant after Benjamini-Hochberg adjustment; the smallest adjusted P value was 0.064. In the reduced seven-predictor sensitivity analysis, AUC was higher than the 52-predictor model in 17 of 26 evaluable subgroup models, with a median descriptive difference of +0.008. Calibration was suboptimal: Across the 26 models, calibration slopes ranged from 0.078 to 0.467 and intercepts from -1.287 to -0.535. Decision-curve analysis showed threshold-dependent net benefit rather than uniform superiority across the evaluated range.

CONCLUSION

Single-lead ECG combined with anthropometric information showed moderate internal discrimination for recorded diabetes status, but the corrected analysis did not identify a cardiac-pathology subgroup effect that remained significant after multiplicity adjustment. The reduced-feature sensitivity analysis suggested that similar discrimination may be achievable with fewer predictors, whereas calibration remained inadequate for direct interpretation of model outputs as disease probabilities. These findings do not establish a clinically validated screening or rule-out threshold. Patient-grouped validation, independent threshold validation, and external validation remain necessary before clinical implementation.

Keywords: Diabetes mellitus; Single-lead electrocardiography; Machine learning; Cardiovascular disease; Model calibration; Decision-curve analysis

Core Tip: This study demonstrates that single-lead electrocardiography combined with anthropometric variables provides moderate internal discrimination of type 1 and type 2 diabetes mellitus, with median areas under the receiver operating characteristic curve of 0.693 and 0.661, respectively. Although model performance varied across cardiac-comorbidity subgroups, none of these differences remained statistically significant after adjustment for multiple comparisons. A reduced seven-predictor model demonstrated broadly comparable discrimination, suggesting that fewer variables may be sufficient. However, substantial probability miscalibration and threshold-dependent clinical net benefit limit immediate clinical applicability. These findings highlight the potential of electrocardiography-based machine learning while emphasizing the need for patient-level and external validation before clinical implementation.