Data from 581 individuals participating in the Diabetes Health Screening Clinic (DiabHealth) at Charles Sturt University, Albury, Australia, were included in the analysis of blood and urine specimens. The study obtained ethical approval from the University Human Ethics Committee under Protocol Number 2006-042, and all participants provided written informed consent, including agreement for the publication of results. A control group with a screening glucose of <5.6 mmol/L, a prediabetes group with a screening glucose of 5.6–6.9 mmol/L, and a type 2 diabetes group with screening glucose ≥7 mm/L were compared. No exclusion criteria were established for this study, as patients participated in a public screening clinic focused on diabetes progression and diabetes complications. Gender, age, and the use of medications, including antiplatelet agents (AntiPL), lipid-lowering agents (AntiLP), antihypertensive medications (HT-med), and antidepressants, were collected for all patients in the study. The ECGs were recorded using a GE Healthcare MAC™ 5500 HD 12-lead electrocardiograph to obtain PQ interval conduction time, QRS duration, QTc interval, and QT dispersion (QTd). These metrics were analyzed from 10-s and 5-min recordings. These intervals represent important electrophysiological mechanisms related to diabetic pathogenesis and cardiovascular consequences. Autonomic dysfunction and decreased parasympathetic regulation in diabetes can affect atrioventricular conduction time, which affects the PQ interval (43, 44). QRS duration measures intraventricular conduction, and small delays may indicate structural or electrical changes like left ventricular hypertrophy or fibrosis (45, 46). The QTc interval and QTd measure ventricular repolarization heterogeneity, which predicts arrhythmic risk and sudden cardiac mortality, especially in diabetics (47–49). In diabetes, extended QTc and elevated QTd are linked to insulin resistance, glycemic management, and cardiovascular mortality (50, 51). These ECG indicators enable a comprehensive assessment of conduction, depolarization, and repolarization issues associated with diabetes and cardiovascular risk. The study focused on ECG biomarkers and diabetes progression, particularly in the context of comorbidities and medication use, over a 10-year period, utilizing data exclusively from first-admission records.

Microsoft Excel (Office 365, Microsoft) was used to analyze descriptive data, and the results are presented as mean ± standard deviation (x ± SD). The machine learning process and statistical analyses were created to determine the significance of characteristics in predicting diabetes status and to find significant group differences. The Synthetic Minority Oversampling Technique (SMOTE), which was used to address class imbalance in the dataset, ensured that the four clinical subgroups—no comorbidities, hypertension, cardiovascular disease, and combined hypertension and cardiovascular disease—all had balanced representation of the no diabetes, prediabetes, and type 2 diabetes mellitus (T2DM) groups (52). The Support Vector Machine (SVM) consistently produced the greatest classification accuracy for differentiating diabetes groups among the machine learning models examined, including logistic regression, random forests, and k-nearest neighbors. It was chosen as the main classifier due to its performance and ability to handle high-dimensional data. SVM maximizes the gap between the closest support vectors (data points closest to the border) by finding the best hyperplane in the feature space that separates data points of various classes (53). This maximum improves the model’s generalizability and reduces the possibility of misclassification. By adjusting the regularization parameter (C), kernel type, and gamma settings, a grid search and 5-fold cross-validation were used to improve the SVM hyperparameters. As a measure of model stability, the associated standard deviation was provided, and the model with the greatest mean cross-validated accuracy was chosen as the best-performing model. The contribution of each feature was then measured by applying permutation significance to the optimum SVM model. This helped determine the most significant predictors by rearranging the values of each characteristic and calculating the change in model accuracy that resulted from these changes. Bar plots were used to display the results, and feature names were reformatted for readability. Pandas, scikit-learn, and imbalanced-learn were used in Python for further statistical studies. Kruskal–Wallis tests for non-parametric group differences were used to evaluate continuous data, whereas Chi-square tests using contingency tables were used to investigate categorical variables (54). To control for the increased risk of Type I errors due to multiple comparisons, we applied the Bonferroni correction by adjusting the significance threshold (α) from the conventional 0.05 level to α/m to include the total number of pairwise tests conducted (55). Significant variables were further investigated using paired Mann–Whitney U tests (56). This integrated methodology provides a comprehensive evaluation of the primary parameters that influence diabetes categorization and development by combining machine learning with robust statistical methodologies.

Data from five hundred and eighty-one participants were included in the analysis.

Table 1 summarizes the demographic and clinical characteristics of the Healthy, Prediabetes, and T2DM participant groups, including pairwise post-hoc Chi-square tests with the Bonferroni correction for categorical variables. Prescription of statins (AntiLP-MedUse) was significantly higher in the T2DM group, with 33.3% of individuals compared to the Healthy and Prediabetes groups (p < 0.001, Kruskal–Wallis). Post-hoc pairwise Chi-square tests with Bonferroni correction revealed a significant difference between the Healthy and T2DM groups (p < 0.001). Medications, such as antidepressant medications, did not vary significantly between groups (p = 0.383). The mean age of 55.22 years in the Healthy group to 58.92 years in the T2DM group was not significant. Similarly, gender distribution was predominantly female in all groups but was not significantly different between the three groups. ECG features, however, reveal a significant difference in QTd-10 s, suggesting a potential link to diabetes. Pairwise comparisons also highlight that QTd-5 min is significantly different between the Healthy and T2DM groups.

Demographic and clinical characteristics of the participants with hypertension are presented in Table 2, along with pairwise post-hoc Chi-square tests with the Bonferroni correction for categorical variables. In Table 2, the term Control refers to patients diagnosed with hypertension only. Medication is a key differentiator among the groups, with AntiLP-MedUse, as indicated by post-hoc pairwise Chi-square tests with Bonferroni correction, showing a significant difference between Healthy and T2DM, and between Prediabetes and T2DM. Diabetes medication was prescribed significantly more often for the hypertension groups compared to the T2DM group. As expected, increased hypertension-specific medication use (HT-MedUse) was observed in all three diabetes-associated groups compared to the Healthy group. ECG features reveal several significant distinctions. PQ-10 s differed significantly between groups, with a significant difference between the Healthy and Prediabetes groups. QTd-10 s and QTd-5 min also show significant differences, particularly between the Prediabetes and T2DM groups.

AntiLP-MedUse in the cardiovascular disease cohort across the three groups was significantly more common in the T2DM group (Table 3). Post-hoc pairwise Chi-square tests with the Bonferroni correction revealed significant differences between the Control and T2DM groups, as well as between the Prediabetes and T2DM groups. In Table 3, the term Control refers to patients diagnosed with cardiovascular disease only. Diabetes medication was only prescribed in the T2DM group. In line with previous results, gender distribution revealed significant variation, with a higher proportion of males in the prediabetes group compared to the control and T2DM groups. Post-hoc pairwise comparisons revealed a significant difference between the Control and Prediabetes groups, but not between Control and T2DM groups, or between Prediabetes and T2DM groups. ECG features show mixed results. PQ-10 s exhibited a significant difference between the prediabetes and T2DM groups. QTc-5 min values were significantly different among the groups, particularly between the control, prediabetes, and T2DM groups, with the longest intervals observed in prediabetes (430.23 ms), which is still within the normal range for both men and women.

ECG biomarkers and medication were able to separate Control, Prediabetes, and T2DM individuals with hypertension (HT) and cardiovascular disease (CVD) (Table 4). In Table 4, the term Control refers to patients diagnosed with both hypertension and cardiovascular diseases. AntiLP-medication use differed significantly amongst diabetes groups, emphasizing its importance in diabetes classification models. Post-hoc comparisons revealed significant differences between the Control and T2DM groups (p < 0.001) and the Prediabetes and T2DM groups (p = 0.005). Previous research has shown that antidiabetic therapy affects disease progression and cardiovascular risk assessment and indicates the importance of using diabetes medication in the model (57). Hypertension medication used in the model may indirectly indicate diabetes progression. T2DM patients also reported higher use of antidepressants compared to Control and Prediabetes (p = 0.010). Since antidepressants have been connected to metabolic abnormalities and diabetes risk, this finding shows a link between depression and diabetes progression that merits additional study.

Several ECG biomarkers showed statistically significant differences, supporting their use in machine learning-based diabetes classification models. QRS-Axis-10 s was significantly different between Control vs. T2DM (p = 0.006), suggesting that diabetic individuals with HT and CVD may have more structural heart or electrical remodeling issues. T2DM patients had longer grade-10 s (p = 0.036) conduction delays, which may indicate autonomic dysfunction and diabetes-related conduction anomalies (58). QTc-10 s (p = 0.001) and QTd-10 s (p = 0.038) demonstrated substantial differences between the three groups, with T2DM patients having wider repolarization times, a risk factor for arrhythmias and sudden cardiac death, specifically between Control and T2DM (p < 0.001) (59).

The feature relevance rankings for diabetes classification in patients without comorbidities using ECG biomarkers over 10 s (Panel A) and 5 min (Panel B) are shown in Figure 1. As the T2DM medication feature is excluded, QRS-Axis-10 s, QTc-10 s, PQ-10 s, and QTd-10 s for 10-s measures and QTd-5 min, QRS-5 min, QRS-Axis-5 min, and PQ-5 min for 5-min measurements are the most influential predictors. Medication-related parameters, especially anti-lipid and AntiLP-MedUse medication use, contribute to classification but are less important.

Combining demographic, clinical, and ECG variables reveals notable differences among the Healthy, Prediabetes, and T2DM groups, as summarized in Table 1. Classification accuracy was 0.86 ± 0.02 for 10-s intervals and 0.88 ± 0.02 for 5-min intervals, with a focus on cardiac ECG biomarkers. ECG biomarkers. Anti-lipid medication use reported by the Healthy and T2DM groups vary significantly. Among ECG biomarkers, QTd-10 s and QTd-5 min show significant differences between the Healthy and Prediabetes groups and between the Healthy and T2DM groups, respectively, as confirmed by pairwise comparisons. In contrast, anti-depression medication does not substantially vary, suggesting that antidepressant use is unrelated to diabetes status in this cohort.

Many factors affect patients with diabetes and hypertension. Without the T2DM medication component, the predictive power shifts primarily to ECG biomarkers. In the 10-s measurement (Panel A of Figure 2), PQ-10 s, QTd-10 s, and QRS-Axis-10 s emerge as the dominant predictors, suggesting their strong role in capturing conduction and repolarization abnormalities. In contrast, the 5-min measurement (Panel B of Figure 2) highlights the most important features, including Birth Age, PQ-5 min, and QRS-Axis-5 min, suggesting that longer ECG recordings capture broader structural and conduction-related differences. The shift in feature rankings between Panels A and B highlights the influence of ECG biomarkers in predicting diabetes risk in hypertensive patients and underscores the role of measurement duration in feature importance distribution. Comparing demographic and clinical factors across the hypertensive diabetic groups shows a significant difference in the reported antilipid medication use between the Control and T2DM groups. Several ECG indicators further distinguish between the three groups. Significant variations in QTd-5 min and QTd-10 s exist between the Prediabetes and T2DM groups, and between the Control and T2DM groups. PQ-10 s difference is significant, with post-hoc comparisons confirming differences between Control and Prediabetes and Prediabetes vs. T2DM. QRS-Axis-10 s also distinguishes Control from T2DM. Demographic characteristics have a to minimal impact on diabetes classification in individuals with hypertension. As for QRS-Axis-5 min, although the overall group difference was marginal (p = 0.055), the difference between Prediabetes and T2DM was significant (p = 0.032), while the other pairwise comparisons were not. The results confirm that ECG biomarkers are key diabetes classification factors for hypertensive patients, with a classification accuracy of 0.72 ± 0.05 for the 10-s recording and 0.74 ± 0.08 for the 5-min recording.

In patients with cardiovascular disease, PQ-10 s, QTc-10 s, QTd-10 s, and QRS-Axis-10 s emerge as the most significant predictors, indicating that short-term conduction and repolarization measures gain prominence (Panel A of Figure 3). In the 5-min measurements, AntiLP-MedUse, QRS-Axis-5 min, and QTc-5 min dominate (Panel B of Figure 3). Notably, PQ-10 s showed marginal overall significance, but its pairwise comparison between Prediabetes and T2DM reached significance (p = 0.019). For the 5-min recordings, AntiLP-Med Use (p = 0.002), QRS-Axis-5 min (p = 0.014), and QTc-5 min (p = 0.048) showed the most substantial group differences, underscoring the dominance of long-term ECG measures and medication use in distinguishing between Control, Prediabetes, and T2DM groups (Panel B, Figure 3). These shifts emphasize the interplay between ECG biomarkers and medication use in cardiovascular disease patients, and the accuracies equal 0.74 ± 0.04 for the 10-s recording and 0.77 ± 0.07 for the 5-min recording. In addition to medication and ECG features being important in classification accuracy, gender indicated significant differences between the diabetes groups. However, the antidepressant medication used does not differ and contributes to the model, demonstrating that antidepressant use does not discriminate between cardiovascular disease patients with diabetes and patients with different comorbidities or no comorbidities.

The classification of diabetes in individuals with hypertension and cardiovascular disease depends primarily on medication use, as expected, but includes ECG biomarkers, with minor contributions from demographic features. In Figure 4A (10-s analysis), QTd-10 s, HT-MedUse, and QRS-Axis-10 s emerge as the top predictors, highlighting the role of repolarization abnormalities and hypertension treatment. In Figure 4B (5-min analysis), AntiLP-MedUse, HT-MedUse, and OtherMeds-Depression dominate, suggesting that non-diabetes medications, particularly lipid-lowering and antidepressant treatments, influence long-term ECG variability. The classification accuracy for these models is 0.68 ± 0.06 in Figure 4A, and 0.64 ± 0.05 in Figure 4B, indicating performance loss. This shows that including provides a focus on physiological indicators. For individuals with hypertension and cardiovascular disease, Table 4 compares demographic, clinical, and ECG data from the Control, Prediabetes, and T2DM groups. HT-MedUse marks a significant difference, with 93.7% of T2DM patients using hypertension medication compared to the Control and Prediabetes groups. AntiLP-MedUse is substantially greater in T2DM patients (67.4%) compared to Prediabetes (34.6%) and Control (41.1%) groups. Post-hoc comparisons reveal substantial differences between the Control and T2DM groups (p < 0.001) and the Prediabetes group. QTd-10 s and QTd-5 min are ECG features with inconsistent results. QTd-10 s is statistically different among groups, whereas QTd-5 min is not. QRS-Axis-10 s differs considerably, especially between Control and T2DM, supporting ECG findings for diabetes categorization. In contrast, age and gender do not significantly affect diabetes classification in this population. OtherMeds-Depression is significantly different, suggesting an association between antidepressant use and diabetes status. Post-hoc comparisons show that Control vs. T2DM shows greater differences, while Prediabetes vs. T2DM does not for anti-depressive medication use.

Shorter ECG recordings improve categorization accuracy, according to the study. Adding the T2DM treatment feature enhances classification accuracy across all groups, boosting biomarker predictive power and model resilience (accuracy up to 0.88 ± 0.04). This suggests that medication data complements, rather than duplicates, diabetes status information. Medication use improved classification accuracy for patients without comorbidities, hypertensive patients, those with cardiovascular disease, and those with both conditions. These findings demonstrate that ECG biomarkers and medication data can be combined to improve the classification of diabetes.

The classification of diabetes status in patients without comorbidities was primarily influenced by QRS-Axis-5 min, QRS-Axis-10 s, QTd-5 min, and QTc-5 min, with a classification accuracy of 0.91 ± 0.05 (Figure 5).

For patients with HT, feature importance rankings shifted, with QRS-Axis-5 min, QTd-10 s, and PQ-10 s becoming the most influential features, and accuracy was 0.79 ± 0.07 (Figure 6).

In patients with CVD, PQ-5 min, Gender, and PQ-10 s emerged as the top predictors (Figure 7), leading to a classification accuracy of 0.82 ± 0.09.

For patients with both HT and CVD, antihypertensive medication use, QTd-10 s, and AntiLP-MedUse were the main features of the model, with a classification accuracy of 0.69 ± 0.07 (Figure 8). More detailed results of the performance metrics, including AUC per class, specificity, and sensitivity, are provided in the Supplementary material.

The distributions of PQ, QRS, QTc, QTd, and QRS-Axis—measured over 10-s and 5-min intervals for all classes (healthy, noDM-HT, noDM-CVD, and noDM-HT + CVD), with and without medication, are presented in Figure 9. Across all features, significant differences were observed between the classes associated with either the 10-s ECG or 5-min ECG recordings, as determined by the Kruskal–Wallis test (p < 0.05). Notably, noDM-HT + CVD patients exhibited prolonged PQ intervals compared to the other groups, reflecting potential cardiac conduction delays in this group (Figure 9A). Healthy patients generally displayed shorter QRS intervals, whereas noDM-HT and noDM-HT + CVD patients showed wider QRS complexes, indicating ventricular conduction abnormalities in these groups (Figure 9B). The QTc features, with QTc-10 s and QTc-5 min, are prolonged in noDM-HT + CVD patients (Figure 9C). This suggests an elevated risk of arrhythmias in patients with combined hypertension and cardiovascular disease, which is consistent with clinical expectations given the known association between QTc prolongation and cardiovascular comorbidities. The dispersion of QT intervals (QTd) was higher in noDM-HT and noDM-HT + CVD groups, particularly the 10-s recordings (Figure 9D). This increased variability could indicate greater heterogeneity in ventricular repolarization in these populations. Finally, significant differences in QRS-Axis-10 s and QRS-Axis-5 min were observed among the control classes (Figure 9E). The noDM-HT and noDM-HT + CVD groups exhibited more negative QRS axes, indicative of left axis deviation, often associated with structural or electrical cardiac abnormalities.

The distributions of PQ, QRS, QTc, QTd, and QRS-Axis, measured over 10-s and 5-min intervals in all prediabetic subclasses (only prediabetes, prediabetes with HT, prediabetes with CVD, and prediabetes with HT + CVD) with and without medication, are presented in Figure 10. Class differences were significant within the same ECG measurement time, as shown by the Kruskal–Wallis test (p < 0.05). The PQ feature distributions show that PQ-10 s and PQ-5 min values differ between classes (Figure 10A). Prediabetes HT and HT + CVD, with medication, had longer PQ intervals than other classes, suggesting atrioventricular conduction delays. QRS features reveal distinct distributional differences across patient groups for both QRS-10 s and QRS-5 min durations (Figure 10B). Notably, the prediabetes groups with hypertension (HT) and with both HT and cardiovascular disease (CVD) exhibit broader and right-shifted distributions, indicating higher median and more variable QRS durations compared to the prediabetes-only group. These findings suggest potential abnormalities in cardiac conduction associated with comorbid HT and CVD. The prediabetes with CVD group is notably broader and skewed toward longer durations, particularly in QRS-5 min recordings. This indicates greater variability and a tendency toward prolonged QRS complexes, supporting the presence of conduction system impairments often associated with cardiovascular disease. Prediabetes CVD, and HT + CVD had longer QTc-10 s and QTc-5 min (Figure 10C), indicating a higher risk of ventricular arrhythmias in this sample. Only the prediabetes group had consistently shorter QTc intervals. QTd features show that prediabetes with HT and prediabetes with HT + CVD had larger QTd dispersion, especially over 10-s intervals, as in Figures 6, 10D. The ventricular repolarization in this group is more heterogeneous, explaining this diversity. QRS-Axis features show class differences over 10-s and 5-min intervals. Prediabetes with HT and prediabetes with HT + CVD groups had greater negative QRS-Axis values, indicating left axis deviation due to structural or electrical cardiac problems (Figure 10E).

PQ, QRS, QTc, QTd, and QRS-Axis feature distributions over 10-s and 5-min intervals in all T2DM classes (T2DM, T2DM with HT, T2DM with CVD, and T2DM with HT + CVD) with and without medication are shown in Figure 11. PQ interval results indicate that PQ-10 s and PQ-5 min periods differ across classes (Figure 11A). T2DM patients with HT and those with HT + CVD had longer PQ intervals than T2DM patients, suggesting atrioventricular conduction delays. Group differences in QRS-10 s and QRS-5 min durations are also present, as illustrated in Figure 11B. T2DM patients had shorter QRS intervals, but T2DM with HT and HT + CVD patients had wider QRS complexes, indicating ventricular conduction problems. QTc-10 s and QTc-5 min intervals are significantly prolonged in T2DM with HT and in patients with HT and CVD (Figure 11C), indicating a higher risk of arrhythmogenic events in this group. In contrast, T2DM patients had shorter QTc intervals. T2DM with CVD and HT + CVD had larger QTd dispersion (Figure 11D). The variability in ventricular repolarization was greatest when the 10-s recoding results were included. QRS-Axis characteristics in Figure 11E indicated substantial group differences in QRS-10 s and QRS-5 min. HT and HT + CVD patients had greater negative QRS-Axis values, indicating structural or electrical cardiac problems associated with left axis deviation.