Machine learning-based detection of diabetes mellitus from single-lead electrocardiography: A phenotype-stratified approach

Abstract

BACKGROUND

Diabetes mellitus (DM) and the related sequalae remains one of the most frequently reported cause of morbidity and mortality in our era. This returns to the non-sufficient screening methods for DM at early stages.

AIM

To assess the diagnostic capabilities of the parameters of single lead electrocardiography (ECG) in the diagnosis of DM utilizing machine learning model.

METHODS

A single center study involved 629 participants with vs without DM. All the study participants passed transthoracic echocardiography, fasting blood glucose measurement, standard 12-lead ECG recording, and single lead ECG registration using the Cardio-Qvark^® device. A gradient boosting machine model, specifically the XGBoost implementation, was developed using R v4.2 and Python v3.10. The model was trained and validated using a novel cluster-stratified approach - training on three phenotypic clusters and testing on the fourth - to isolate DM-specific ECG signatures from confounding cardiovascular disease.

RESULTS

The cluster-stratified analysis revealed that the model performed best in cluster 4 (patients with high DM prevalence and significant comorbidities), achieving a sensitivity of 75%, specificity of 83%, and an area under the curve of 88%.

CONCLUSION

This study demonstrates that a phenotype-stratified approach is crucial for effective ECG-based DM screening. By identifying a specific clinical profile (cluster 4: High comorbidity burden with preserved cardiac function), we developed a model that accurately detects DM from a single-lead ECG. This phenotype-specific strategy overcomes the confounding effect of cardiovascular disease, moving beyond one-size-fits-all algorithms towards a precise and clinically viable tool for non-invasive DM detection in high-risk populations.

Key Words: Diabetes mellitus; Machine learning model; Diagnosis; Single lead electrocardiography; Hyperglycemia; Metabolic syndrome

Core Tip: This study introduces a crucial paradigm shift for non-invasive diabetes detection using electrocardiography. Instead of a one-size-fits-all model, we employed phenotypic clustering to disentangle the confounding effects of cardiovascular disease. We identified a specific patient profile (cluster 4: High diabetes prevalence with significant but non-severe comorbidities) where a machine learning model, analyzing single-lead electrocardiography features like T-wave morphology and atrial conduction, achieves optimal and clinically viable performance (area under the curve: 0.88). This proves that diabetes-specific cardiac “whispers” are detectable, but only with a precision medicine approach that tailors diagnostics to distinct clinical phenotypes.

INTRODUCTION

Diabetes mellitus (DM) represents one of the most significant global public health challenges of the 21^st century, with prevalence rates continuing to rise alarmingly[1]. According to the International Diabetes Federation, approximately 589 million adults were living with diabetes in 2024, a figure projected to reach 853 million by 2050, disproportionately affecting low- and middle-income countries[2]. This epidemic carries devastating human and economic costs, driving increased morbidity and mortality through microvascular complications (retinopathy, nephropathy, neuropathy) and macrovascular sequelae [cardiovascular disease (CVD), stroke, peripheral artery disease][3-5]. Critically, an estimated 44% of diabetes cases remain undiagnosed globally, representing a substantial missed opportunity for early intervention that could prevent or delay debilitating complications[6]. The current diagnostic paradigm relies on invasive, resource-intensive, and often inaccessible methods - fasting plasma glucose, oral glucose tolerance tests, and glycated hemoglobin (HbA1c) - which suffer from limitations including cost, requirement for venous blood sampling, limited availability in resource-poor settings, and variable reliability (e.g., HbA1c influenced by hemoglobinopathies or anemia). Consequently, there is an urgent, unmet need for novel, non-invasive, scalable, and cost-effective screening tools capable of facilitating widespread early detection, particularly in underserved populations and primary care settings.

Electrocardiography (ECG), a ubiquitous, non-invasive, and low-cost technology, has emerged as a promising candidate for such a tool. The convergence of single-lead ECG devices and sophisticated machine learning (ML) represents a particularly novel and transformative approach. While chronic hyperglycemia in DM induces subtle alterations in cardiac electrophysiology (e.g., in autonomic tone, repolarization, and conduction)[7-10], these changes are often multidimensional and imperceptible to conventional analysis. ML algorithms are uniquely suited to decipher this complex, subclinical signature from a simple single-lead recording. Furthermore, the advent of portable, smartphone-compatible single-lead ECG devices eliminates the need for complex 12-lead setups, creating an unprecedented opportunity for scalable, accessible screening. This combination of accessible hardware and advanced analytics forms the core of our innovative strategy, moving beyond traditional ECG parameters to a high-dimensional, ML-driven feature extraction for DM detection.

However, translating the potential of ECG-based DM detection into clinical reality faces a fundamental methodological challenge: The confounding influence of prevalent CVD. Diabetes rarely exists in isolation; it is a potent risk factor for and frequently coexists with conditions such as hypertension, coronary artery disease, heart failure, and arrhythmias like atrial fibrillation[11-13]. These CVD states themselves induce profound changes in the ECG signal, often overshadowing or mimicking the subtler alterations potentially attributable to DM per se. Consequently, models trained to detect DM using ECG features in mixed populations risk identifying underlying CVD rather than diabetes specifically, acting as a surrogate marker for a correlated outcome rather than the target pathology. This confounding effect significantly undermines the specificity and clinical utility of proposed ECG-based DM diagnostic algorithms and represents a critical barrier to their real-world implementation. Previous studies attempting ECG-based DM detection have often inadequately addressed this confounder, typically employing simple case-control designs (DM+ vs DM-) without sufficiently accounting for the heterogeneity and burden of underlying CVD within and between these groups.

Therefore, to advance the field beyond this significant limitation, a more nuanced approach is required - one that explicitly disentangles the electrophysiological signature of DM from the background noise of prevalent CVD. Phenotypic stratification, based on comprehensive clinical profiling, offers a potential solution. By identifying homogeneous patient subgroups (clusters) with distinct comorbidity burdens and cardiovascular phenotypes, it becomes possible to investigate whether DM imparts a detectable and distinct ECG signature within specific clinical contexts. This approach acknowledges the complex interplay between DM and CVD and seeks to define the conditions under which DM-related ECG changes are most discernible. Furthermore, developing diagnostic models tailored to these specific phenotypic clusters holds the promise of significantly improved accuracy and generalizability compared to generic, population-wide algorithms.

This study aimed to overcome the critical confounding influence of CVD on ECG-based DM detection by employing phenotypic clustering. Specifically, we sought to: (1) Identify distinct clinical phenotypes within a diverse patient cohort using comprehensive cardiovascular and demographic characteristics; (2) Determine the cluster(s) where DM exerts the most discernible influence on the single-lead ECG signal; and (3) Develop and validate cluster-specific ML models optimized for detecting DM within the most phenotypically relevant subgroup, thereby providing a more accurate and clinically applicable approach for non-invasive DM screening. Therefore, our study considered the first globally that apply phenotypic stratification to single-lead ECG for DM detection.

MATERIALS AND METHODS

General study characteristics

A prospective observational non-randomized cross-sectional study was conducted at the I.M. Sechenov First Moscow State Medical University (Sechenov University), Moscow, Russia. The study protocol was approved by the Local Ethical Committee of Sechenov University (approval No. 19-23). The study was also registered on the ClinicalTrials.gov website (ID: NCT04788342). An informed written consent obtained from the study participants.

The study comprised two stages: The first stage involved the consecutive enrollment of 200 patients into the main sample according to the inclusion criteria specified below and the development of an algorithm for detecting DM using single-lead ECG. The second stage involved the additional enrollment of 100 patients into the test sample using the same inclusion criteria as the first stage, for the purpose of analyzing the effectiveness of the developed algorithm. A total of 629 patients’ parameters were included in the study. The proportion of patients with DM in both the training and test samples was required to exceed 30%.

Patient inclusion, and exclusion criteria

Patients from University Clinical Hospital number 2 undergoing outpatient or inpatient examination and treatment, including in the endocrinology department with a diagnosis of DM, were included in the study. All patients reviewed and signed written informed consent to participate. The inclusion and exclusion criteria are presented in the Table 1.

Table 1 Patient inclusion and exclusion criteria for the study.

Criterion type	Description
Inclusion	Patient age over 18 years; agreement to participate in the study
Exclusion	Significant QRS complex morphology changes (such as bundle branch block and ventricular extrasystole); poor quality electrocardiography recorded from the fingers (Parkinson’s disease, tremor of any origin, mental disorders). Unsatisfactory quality of electrocardiography and/or photoplethysmography; withdrawal of consent for further participation in the study

Comprehensive examination

Clinical examination: Collection of complaints, auscultation, palpation, percussion, assessment of edema syndrome, measurement of heart rate and blood pressure level. Medical history analysis: Collection of cardiological history and DM history. The study adjusted medication use, including beta-blockers, angiotensin-converting enzyme inhibitors, sodium-glucose cotransporter 2 inhibitors, that could easily be responsible for the observed ECG changes.

Transthoracic echocardiography: Performed using a GE Vivid 5 device. Assessment of cardiac chamber volumes, wall thickness, diastolic and systolic function status, and detection of valvular apparatus and major cardiac vessel dysfunction according to current guidelines[14]. Coronary artery disease diagnosis obtained from the anamneses of the patient and the drugs that take for it is management.

Standard 12-lead ECG recording: Performed using a SCHILLER AT-5 device. Determination of temporal and amplitude ECG parameters, cardiac electrical axis direction, and assessment of cardiac cycle morphology.

Capillary blood glucose analysis: Fasting blood glucose level measured twice within 1-3 days, in a seated position, adhering to the following requirements: Analysis performed after a 12-hour fast; abstinence from alcohol 2-3 days prior; avoidance of stress or heavy physical exertion; refraining from tooth brushing, smoking, or consumption of fatty foods, sugary drinks, or confectionery products the day before. The diagnostic thresholds for DM are a fasting plasma glucose of 126 mg/dL (7 mmol/L) or higher and the presence of DM history. The type of the DM is classified based on the history of the patients, and the drug they take to it is management.

Single-lead ECG and photoplethysmography recording: Recorded using the Qardio-Qvark^® device (LLC “L Card”, Moscow, Russia; registered with the Federal Service for Surveillance in Healthcare No. RZN 2019/8124 dated February 15, 2019) immediately after each blood glucose analysis. Recordings lasted 1 minute, performed at rest in a seated position in a blind manner. The recorder has a smartphone case form factor and features 2 sensors for recording ECG and photoplethysmography. Recordings were transmitted to a cloud-based platform for analysis using proprietary ML algorithms.

The Qardio-Qvark^® embedded algorithms perform mathematical calculations obtained from the single lead ECG (lead I). Actually, these calculations are mathematical formulas and of importance to the mathematicians who build this algorithm. The algorithm shows 47 parameters that based on some mathematical features obtained from the single lead ECG, can be found on our previous publication[15]. The algorithm also, can predict, some disease, but further investigations are done to enhance its diagnostic accuracy[15].

Statistical analysis

Statistical analysis was performed using the R programming language (v4.3). For quantitative variables, the distribution type (using the Shapiro-Wilk test), mean, standard deviation, median, interquartile range (IQR), minimum, and maximum values were determined. For categorical and qualitative variables, the proportion and absolute number of values were calculated.

Comparative analysis for normally distributed quantitative variables was performed using Welch’s t-test (for 2 groups) or analysis of variance (for more than 2 groups) with subsequent pairwise group comparisons; for non-normally distributed quantitative variables, the Mann-Whitney U test (for 2 groups) or the Kruskal-Wallis test (for more than 2 groups) was used. Comparative analysis of categorical and qualitative variables was performed using Pearson’s χ² test, and, when not applicable, Fisher’s exact test. When adjusting for multiple comparisons, the Holm method was used.

Diabetes diagnosis using single-lead ECG data

Since MD is a background condition for many CVDs, there may be an issue of correlation between the presence of diabetes and changes observed on the ECG. To reduce the influence of confounding factors, the clinical and demographic data were clustered (using Ward’s method, with prior normalization of quantitative variables and binarization of categorical variables). These data were not used in model training, but were used to determine the corresponding cardiovascular phenotype of each individual. Firstly, this approach was aimed at identifying clusters of individuals without CVDs, or with mild stages of such diseases. Secondly, the prevalence of diabetes in each cluster was assessed.

The clusters were then used to attempt to build descriptive models based on sex, age, smoking status, rhythm disturbances during recording, and the single-lead ECG data itself. For validation, a specific cluster would be used, while all others would be used for training. In this way, it would be possible to test the feasibility of diagnosing diabetes among different cardiovascular phenotypes, as well as to identify factors with high importance that recur in two or more models.

In other words, validation would be performed for different phenotypes, and technically, instead of typical cross-validation, validation on clusters would be used. The training pipeline included normalization of quantitative variables and binarization of categorical variables. Model performance was evaluated on the validation cluster using area under the curve (AUC), sensitivity, specificity, positive and negative predictive values (with the optimal threshold determined by the Youden index on the training data). The input features intent to cluster data includes clinical features and the 47 features of the single lead ECG. The optimal number of clusters was determined by evaluating the dendrogram structure and the average silhouette width, with the goal of maximizing clinical interpretability and internal cohesion.

Prior to model training, a structured preprocessing pipeline was applied to the input features, which included both clinical variables and the 47 single-lead ECG parameters. Missing data were first assessed; variables with more than 5% missing values were excluded from the analysis, while those with fewer missing values were imputed using the median for continuous variables and the mode for categorical variables. Outliers in continuous ECG parameters were identified using the IQR method, where values less than Q1 - 1.5 × IQR or greater than Q3 + 1.5 × IQR were Winsorized (i.e., capped at the 5^th and 95^th percentiles) to reduce their influence without discarding data. Finally, all continuous variables were normalized to a mean of zero and a standard deviation of one to ensure features were on a comparable scale for the gradient boosting algorithm.

The 47 parameters derived from the single-lead ECG signal encompass a range of temporal, amplitude, and morphological features. For clarity, the physiological basis and calculation of key parameters discussed in the results are defined here: T-wave flatness index (Tfi): A quantitative measure of T-wave morphology, calculated as the ratio of the T-wave’s area to its amplitude. A lower value indicates a flatter, more rounded T-wave, which is associated with altered ventricular repolarization heterogeneity. Corrected QT (QTc) interval: The QT interval adjusted for heart rate using Bazett’s formula (QTc = QT/), representing the duration of ventricular repolarization. QRS energy in band 4 (QRSE4): The spectral energy of the QRS complex within a specific high-frequency band (typically 40-100 Hz), reflecting the integrity and velocity of ventricular depolarization. High-frequency QRS (HFQRS): The root mean square voltage of the QRS complex in a high-frequency band (typically 150-250 Hz), sensitive to changes in myocardial conduction properties. Ventricular activation time (VAT): The interval from the onset of the QRS complex to the peak of the R-wave, representing the initial depolarization time of the ventricles. P-wave start time (Pst) and P-wave finish time: The onset and offset timings of the P-wave, defining the total duration of atrial depolarization.

Model training and validation strategy

A nested, cluster-stratified validation framework was employed to ensure robust model evaluation and mitigate overfitting. The overall dataset was first divided into the four predefined phenotypic clusters. For each iteration of the outer validation loop, one entire cluster was held out as the validation set. The remaining three clusters were combined to form the training set.

Within this training set, a 5-fold cross-validation process was implemented for hyperparameter tuning of the gradient boosting machine model. The hyperparameters yielding the best average performance across these 5 folds were then used to retrain the model on the entire training set (all three clusters). This final model was subsequently evaluated on the held-out validation cluster to estimate its generalizable performance on a distinct phenotypic group. This process was repeated four times, with each cluster serving as the validation set once, and the performance metrics were reported for each validation cluster individually.

RESULTS

The comparative features of the two groups of participants, with and without DM represented in the table below (Table 2). Due to the previously identified target variable likely reflects ECG alterations associated with underlying CVD, with DM serving as a comorbid factor (though this requires verification, which is methodologically challenging). Effectively, a correlated surrogate outcome - distinct from our hypothesized target - was measured.

Table 2 Comparative statistics of the groups, n (%)/mean ± SD.

Factor	Index	Without DM	With DM	Р value
Age (years)		56.1 ± 17.1	63.7 ± 10.6	0.000
Sex	Female	177 (38.6)	83 (48.5)	0.032
Left ventricle diastolic dysfunction	No	333 (72.7)	74 (43.3)	0.000
	Yes	125 (27.3)	97 (56.7)
Left ventricle diastolic dysfunction_1 stage	No	340 (74.2)	84 (49.1)	0.000
	Yes	118 (25.8)	87 (50.9)
Left ventricle diastolic dysfunction 2_3 stage	No	418 (91.3)	132 (77.2)	0.000
	Yes	40 (8.7)	39 (22.8)
Hypertension degree	No	190 (41.5)	40 (23.4)	0.000
	1st degree	39 (8.5)	25 (14.6)
	2nd degree	140 (30.6)	50 (29.2)
	3rd degree	89 (19.4)	56 (32.7)
Aterial fibrillation at the registration moment (from the single lead ECG)	No	435 (95.0)	162 (94.7)	1.000
	Yes	23 (5.0)	9 (5.3)
Ischemic heart disease	No	361 (78.8)	113 (66.1)	0.001
	1st functional class	68 (14.8)	48 (28.1)
	2nd functional class	29 (6.3)	10 (5.8)
Disturbance of cardiac rhythm and conduction	No	443 (96.7)	162 (94.7)	0.092
	1	11 (2.4)	3 (1.8)
	2	3 (0.7)	5 (2.9)
	3	1 (0.2)	1 (0.6)
Chronic heart failure	No	407 (89.6)	150 (88.2)	0.324
	1st stage	7 (1.5)	0 (0.0)
	2nd stage	22 (4.8)	10 (5.9)
	3rd stage	17 (3.7)	9 (5.3)
	4th stage	1 (0.2)	1 (0.6)
Aortic valve insufficiency	No	351 (76.6)	147 (86.0)	0.022
	1st stage	95 (20.7)	23 (13.5)
	2nd stage	12 (2.6)	1 (0.6)
Aortic valve stenosis	No	443 (96.7)	163 (95.9)	0.753
	1st stage	9 (2.0)	5 (2.9)
	2nd stage	4 (0.9)	2 (1.2)
	3rd stage	2 (0.4)	0 (0.0)
Mitral valve insufficiency	No	134 (29.5)	105 (61.4)	0.000
	1st stage	274 (60.4)	60 (35.1)
	2nd stage	39 (8.6)	6 (3.5)
	3rd stage	7 (1.5)	0 (0.0)
Mitral valve stenosis	No	454 (99.1)	166 (97.1)	0.053
	1st stage	3 (0.7)	5 (2.9)
	2nd stage	1 (0.2)	0 (0.0)
Ejection fraction		59.5 ± 8.6	58.1 ± 8.0	0.005
Ejection fraction > 55	No	73 (15.9)	42 (24.6)	0.018
	Yes	385 (84.1)	129 (75.4)
Ejection fraction < 40	No	436 (95.2)	164 (95.9)	0.870
	Yes	22 (4.8)	7 (4.1)
Ejection fraction < 30	No	453 (98.9)	169 (98.8)	1.000
	Yes	5 (1.1)	2 (1.2)
End diastolic volume (mL)		93.8 ± 36.2	99.9 ± 46.7	0.632
Left atrial volume index		34.9 ± 11.1	36.7 ± 8.1	0.000
Smoking	No	371 (81.0)	154 (90.1)	0.009
	Yes	87 (19.0)	17 (9.9)
SDLA		11.0 ± 5.2	10.4 ± 2.7	0.176
HFNoise		60.7 ± 28.3	66.0 ± 25.5	0.003
RR		822.0 ± 156.2	811.2 ± 141.7	0.533
TpTe		89.4 ± 20.3	94.3 ± 29.7	0.487
VAT		38.8 ± 11.1	41.6 ± 13.4	0.020
QTc		425.8 ± 44.4	447.0 ± 52.4	0.000
QT_TQ		1.0 ± 0.4	1.0 ± 0.4	0.001
HFQRS		15919.9 ± 199679.8	10.3 ± 8.5	0.989
HFSNR		2.5 ± 2.3	2.1 ± 1.8	0.149
JA		-24.9 ± 57.7	-35.9 ± 81.8	0.054
J80A		1504.0 ± 31942.7	-3.8 ± 92.9	0.025
TA		185.0 ± 153.9	151.2 ± 162.2	0.006
QRSenergy		616.1 ± 557.9	728.6 ± 870.9	0.117
Tenergy		108.4 ± 101.3	111.7 ± 113.7	0.926
Tpenergy		77.0 ± 86.1	73.6 ± 89.1	0.218
Sbeta		0.9 ± 0.5	1.0 ± 0.9	0.003
Beta		0.8 ± 0.4	0.9 ± 0.5	0.000
QRS11energy		539.0 ± 692.3	713.6 ± 1339.7	0.075
QRS12energy		809.2 ± 1129.4	1074.5 ± 1901.7	0.077
QRS2energy		226.5 ± 301.4	305.8 ± 551.4	0.042
QRSE1		4730.9 ± 94779.9	2707.4 ± 34733.0	0.004
QRSE2		1303.9 ± 25310.4	164.5 ± 290.3	0.040
QRSE3		224.6 ± 202.5	266.1 ± 323.5	0.177
QRSE4		230.8 ± 189.8	238.7 ± 194.6	0.454
TE1		45.2 ± 57.1	54.0 ± 72.5	0.263
TE2		2548.4 ± 53784.0	31.7 ± 27.2	0.186
TE3		3621.5 ± 77261.8	9.5 ± 9.2	0.019
TE4		10275.1 ± 134000.1	5.2 ± 5.1	0.031
QRSw		88.7 ± 23.5	92.3 ± 25.9	0.078
PAn		67.3 ± 37.6	60.5 ± 40.1	0.081
Pan_1		-1.4 ± 17.6	-1.6 ± 10.3	0.981
RA		903.0 ± 414.2	916.9 ± 432.6	0.794
SA		180.0 ± 155.1	198.7 ± 209.0	0.632
Pst		217.6 ± 121.4	144.2 ± 123.3	0.000
Pfi		77.3 ± 77.0	101.5 ± 79.9	0.000
QRSst		165.6 ± 85.6	190.7 ± 88.4	0.000
QRSfi		436.2 ± 117.0	371.1 ± 169.8	0.000
Tfi		5.5 ± 170.3	293.1 ± 301.6	0.000
PpeakP		-67.5 ± 55.2	-7.2 ± 71.5	0.000
PpeakP		107.9 ± 81.7	80.7 ± 106.0	0.000
Rpeak		138.4 ± 92.7	197.8 ± 94.6	0.000
Speak		176.0 ± 109.5	223.3 ± 102.9	0.000
Tpeak		375.6 ± 106.9	453.2 ± 123.7	0.000
Tons		329.7 ± 105.4	403.1 ± 118.7	0.000
Toffs		416.8 ± 113.5	493.4 ± 126.5	0.000
RonsF		34.1 ± 4.7	33.0 ± 5.8	0.030
RoffsF		33.2 ± 6.0	32.1 ± 6.8	0.078
SDNN		2776.7 ± 58807.6	25.9 ± 19.1	0.030

The used statistical tests include Mann-Whitney U test, χ² test, and Fisher’s exact test as per demand. Starting from the standard deviation of all normal-to-normal intervals to the standard deviation of normal-to-normal intervals, this part of the table represents the parameters obtained from the Qardio-Qvark^® device based on some mathematical formulas. ECG: Electrocardiography; SDLA: Standard deviation of all normal-to-normal intervals; HFNoise: High frequency noise; RR: RR interval; TpTe: Tpeak-tend; VAT: Ventricular activation time; QTc: Corrected QT; HFQRS: High-frequency QRS; HFSNR: High frequency signal-to-noise ratio; JA: J-point amplitude; J80A: J-point + 80 milliseconds, amplitude; TA: T-wave amplitude; QRSE: QRS energy/spectral integral; TE: T-wave energy; QRSw: QRS width; PAn: P-wave amplitude negative; RA: R-wave amplitude; SA: S-wave amplitude; Pst: P-wave start; Pfi: P-wave finish; QRSst: QRS start; QRSfi: QRS finish/end; Tfi: T-wave flatness index; PpeakP: P-peak amplitude; Rpeak: R-wave amplitude; Speak: S-peak amplitude; Tpeak: T-peak amplitude; RonsF: R onset; RoffsF: R offset; SDNN: Standard deviation of normal-to-normal intervals.

To address this issue a methodological framework performed through the identification of phenotypic subgroups stratified by DM status. Were the entire dataset was clustered using the following baseline clinical and demographic variables - representing all available parameters (notably limited in scope): Age, sex, diastolic dysfunction, hypertension, atrial fibrillation, coronary artery disease, conduction disorders, chronic heart failure, valve stenosis/insufficiency, ejection fraction, end-diastolic volume, obesity, DM, smoking status, and pulmonary artery systolic pressure.

Clustering outcome

Four distinct clusters emerged: Cluster 1: Younger, predominantly healthy individuals. Cluster 2: Older, predominantly healthy individuals. Clusters 3 and 4: Patients with significant comorbidities. These clusters define four phenotypic patient profiles. The average silhouette width for this 4-cluster solution was 0.32, indicating a reasonable and meaningful structure in the data (Figure 1). Comparative analyses and pairwise statistics are detailed in the below table, for full inter-cluster analysis in Table 3 and Supplementary Table 1.

Figure 1 Phenotypic clusters identified by clinical profiling. The plot illustrates the four distinct patient clusters derived from hierarchical clustering of clinical and demographic variables. Cluster 1 younger, predominantly healthy individuals. Cluster 2 older, predominantly healthy individuals. Clusters 3 and 4: Patients with significant comorbidities but divergent profiles.

Table 3 Comparative descriptive statistics of the clusters, n (%)/mean ± SD.

Factor	Statistic	Cluster 1	Cluster 2	Cluster 3	Cluster 4	P value
Age (years)		43.1 ± 16.2	62.1 ± 12.6	64.8 ± 12.2	64.8 ± 10.3	0.000
Sex	Female	72 (44.2)	112 (48.3)	53 (37.3)	23 (25.0)	0.001
	Male	91 (55.8)	120 (51.7)	89 (62.7)	69 (75.0)
Left ventricle diastolic dysfunction	No	146 (89.6)	166 (71.6)	51 (35.9)	44 (47.8)	0.000
	Yes	17 (10.4)	66 (28.4)	91 (64.1)	48 (52.2)
Atrial fibrillation at the time of single lead ECG registration	No	163 (100.0)	226 (97.4)	119 (83.8)	89 (96.7)	0.000
	Yes	0 (0.0)	6 (2.6)	23 (16.2)	3 (3.3)
Ejection fraction (%)		61.4 ± 5.7	61.5 ± 5.1	52.9 ± 12.1	58.7 ± 7.7	0.000
End diastolic volume (mL)		81.7 ± 24.4	87.7 ± 24.0	113.1 ± 51.4	101.0 ± 42.3	0.000
Left atrial volume index		30.3 ± 5.6	34.4 ± 7.8	42.3 ± 15.3	36.1 ± 6.9	0.000
Diabetes mellitus	No	131 (80.4)	163 (70.3)	108 (76.1)	56 (60.9)	0.005
	Yes	32 (19.6)	69 (29.7)	34 (23.9)	36 (39.1)
Smoking	No	137 (84.0)	200 (86.2)	114 (80.3)	74 (80.4)	0.398
	Yes	26 (16.0)	32 (13.8)	28 (19.7)	18 (19.6)
Pulmonary artery systolic pressure		10.0 ± 0.0	10.0 ± 0.0	13.6 ± 9.3	10.0 ± 0.0	0.000
Hypertension_0	No	0 (0.0)	211 (90.9)	113 (79.6)	75 (81.5)	0.000
	Yes	163 (100.0)	21 (9.1)	29 (20.4)	17 (18.5)
Hypertension_1	No	163 (100.0)	178 (76.7)	138 (97.2)	86 (93.5)	0.000
	Yes	0 (0.0)	54 (23.3)	4 (2.8)	6 (6.5)
Hypertension_2	No	163 (100.0)	100 (43.1)	102 (71.8)	74 (80.4)	0.000
	Yes	0 (0.0)	132 (56.9)	40 (28.2)	18 (19.6)
Hypertension_3	No	163 (100.0)	207 (89.2)	73 (51.4)	41 (44.6)	0.000
	Yes	0 (0.0)	25 (10.8)	69 (48.6)	51 (55.4)
Ischemic heart disease_0	No	0 (0.0)	1 (0.4)	63 (44.4)	91 (98.9)	0.000
	Yes	163 (100.0)	231 (99.6)	79 (55.6)	1 (1.1)
Ischemic heart disease_1	No	163 (100.0)	231 (99.6)	93 (65.5)	26 (28.3)	0.000
	Yes	0 (0.0)	1 (0.4)	49 (34.5)	66 (71.7)
Ischemic heart disease_2	No	163 (100.0)	232 (100.0)	128 (90.1)	67 (72.8)	0.000
	Yes	0 (0.0)	0 (0.0)	14 (9.9)	25 (27.2)
Disturbance of cardiac rhythm and conduction_0	No	0 (0.0)	0 (0.0)	24 (16.9)	0 (0.0)	0.000
	Yes	163 (100.0)	232 (100.0)	118 (83.1)	92 (100.0)
Disturbance of cardiac rhythm and conduction_1	No	163 (100.0)	232 (100.0)	128 (90.1)	92 (100.0)	0.000
	Yes	0 (0.0)	0 (0.0)	14 (9.9)	0 (0.0)
Disturbance of cardiac rhythm and conduction_2	No	163 (100.0)	232 (100.0)	134 (94.4)	92 (100.0)	0.000
	Yes	0 (0.0)	0 (0.0)	8 (5.6)	0 (0.0)
Disturbance of cardiac rhythm and conduction_3	No	163 (100.0)	232 (100.0)	140 (98.6)	92 (100.0)	0.069
	Yes	0 (0.0)	0 (0.0)	2 (1.4)	0 (0.0)
Chronic heart failure_0	No	0 (0.0)	0 (0.0)	67 (48.6)	0 (0.0)	0.000
	Yes	163 (100.0)	231 (100.0)	71 (51.4)	92 (100.0)
Chronic heart failure_1	No	163 (100.0)	231 (100.0)	131 (94.9)	92 (100.0)	0.000
	Yes	0 (0.0)	0 (0.0)	7 (5.1)	0 (0.0)
Chronic heart failure_2	No	163 (100.0)	231 (100.0)	106 (76.8)	92 (100.0)	0.000
	Yes	0 (0.0)	0 (0.0)	32 (23.2)	0 (0.0)
Chronic heart failure_3	No	163 (100.0)	231 (100.0)	112 (81.2)	92 (100.0)	0.000
	Yes	0 (0.0)	0 (0.0)	26 (18.8)	0 (0.0)
Chronic heart failure_4	No	163 (100.0)	231 (100.0)	136 (98.6)	92 (100.0)	0.063
	Yes	0 (0.0)	0 (0.0)	2 (1.4)	0 (0.0)
Aortic valve stenosis_0	No	0 (0.0)	55 (23.7)	59 (41.5)	17 (18.5)	0.000
	Yes	163 (100.0)	177 (76.3)	83 (58.5)	75 (81.5)
Aortic valve insufficiency_1	No	163 (100.0)	177 (76.3)	96 (67.6)	75 (81.5)	0.000
	Yes	0 (0.0)	55 (23.7)	46 (32.4)	17 (18.5)
Aortic valve insufficiency_2	No	163 (100.0)	232 (100.0)	129 (90.8)	92 (100.0)	0.000
	Yes	0 (0.0)	0 (0.0)	13 (9.2)	0 (0.0)
Aortic valve stenosis_0	No	0 (0.0)	0 (0.0)	22 (15.6)	0 (0.0)	0.000
	Yes	163 (100.0)	232 (100.0)	119 (84.4)	92 (100.0)
Aortic valve stenosis_1	No	163 (100.0)	232 (100.0)	127 (90.1)	92 (100.0)	0.000
	Yes	0 (0.0)	0 (0.0)	14 (9.9)	0 (0.0)
Aortic valve stenosis_2	No	163 (100.0)	232 (100.0)	135 (95.7)	92 (100.0)	0.000
	Yes	0 (0.0)	0 (0.0)	6 (4.3)	0 (0.0)
Aortic valve stenosis_3	No	163 (100.0)	232 (100.0)	139 (98.6)	92 (100.0)	0.078
	Yes	0 (0.0)	0 (0.0)	2 (1.4)	0 (0.0)
Mitral valve insufficiency_0	No	59 (36.2)	145 (62.5)	121 (87.7)	61 (66.3)	0.000
	Yes	104 (63.8)	87 (37.5)	17 (12.3)	31 (33.7)
Mitral valve insufficiency_1	No	104 (63.8)	102 (44.0)	52 (37.7)	33 (35.9)	0.000
	Yes	59 (36.2)	130 (56.0)	86 (62.3)	59 (64.1)
Mitral valve insufficiency_2	No	163 (100.0)	217 (93.5)	110 (79.7)	90 (97.8)	0.000
	Yes	0 (0.0)	15 (6.5)	28 (20.3)	2 (2.2)
Mitral valve insufficiency_3	No	163 (100.0)	232 (100.0)	131 (94.9)	92 (100.0)	0.000
	Yes	0 (0.0)	0 (0.0)	7 (5.1)	0 (0.0)
Mitral valve stenosis_0	No	0 (0.0)	0 (0.0)	9 (6.3)	0 (0.0)	0.000
	Yes	163 (100.0)	232 (100.0)	133 (93.7)	92 (100.0)
Mitral valve stenosis_1	No	163 (100.0)	232 (100.0)	134 (94.4)	92 (100.0)	0.000
	Yes	0 (0.0)	0 (0.0)	8 (5.6)	0 (0.0)
Mitral valve stenosis_2	No	163 (100.0)	232 (100.0)	141 (99.3)	92 (100.0)	0.356
	Yes	0 (0.0)	0 (0.0)	1 (0.7)	0 (0.0)

The numbers associated with the name of the diseases represent the stage of the disease. 0: Absent, 1: 1^st stage, 2: 2^nd stage, 3: 3^rd stage, 4: 4^th stage. ECG: Electrocardiography.

Cluster-specific model development and validation

Four gradient boosting models were trained using cluster-stratified cross-validation: Training: Three clusters. Validation: The rest cluster. Features: Age, sex, atrial fibrillation at recording, smoking status, single-lead ECG parameters. The models performance metrics represented in Table 4. Key observations include optimal DM detection occurred in cluster 1 (younger/healthy cohort), cluster 3 exhibited the poorest performance, and the cluster 4 model was selected as the most balanced (Figure 1). The top 5 important features for each cluster with the importance coefficient in the diagnosis of DM represented below, for full feature importance in Table 5 and Supplementary Tables 2-5.

Table 4 The models performance metrics.

Validation cluster	Area under the receiver operating characteristic curve	Sensitivity	Specificity	Positive predictive value	Negative predictive value	Threshold
Cluster 1	0.947	0.969	0.781	0.948	0.862	0.350
Cluster 2	0.835	0.816	0.725	0.875	0.625	0.221
Cluster 3	0.589	0.398	0.941	0.956	0.330	0.131
Cluster 4	0.880	0.750	0.833	0.875	0.682	0.164

Table 5 The top 5 important features for each cluster with the importance coefficient.

Cluster 1; feature (importance coefficient)	Cluster 2; feature (importance coefficient)	Cluster 3; feature (importance coefficient)	Cluster 4; feature (importance coefficient)
Tfi (0.203)	QRSfi (0.206)	Tfi (0.593)	Tfi (0.235)
HFQRS (0.043)	Age (0.120)	QTc (0.053)	Age (0.076)
RR (0.041)	Toffs (0.106)	Age (0.049)	PpeakN (0.051)
Beta (0.033)	QRSE1 (0.053)	Sbeta (0.020)	Rpeak (0.043)
QRSE4 (0.032)	QTc (0.040)	QRSE4 (0.019)	Pst (0.035)

Tfi: T-wave flatness index; QRSfi: QRS complex morphology; HFQRS: High-frequency QRS; QTc: Corrected QT; RR: RR interval; PpeakN: Negative P-wave peak amplitude; QRSE: QRS energy/spectral integral; Rpeak: R-wave amplitude; Pst: P-wave start time.

Tfi reflecting impaired ventricular repolarization, is the top feature in clusters 1 (young/healthy), 3 (severe CVD), and 4 (high-DM). Its importance is high in clusters 1 and 4 where DM drives changes, but low in cluster 3 where CVD dominates (Figure 2). Age highlights age-dependent remodeling and is important in cluster 2 (older/healthy), but less relevant in clusters 3 and 4 due to comorbidities. QTc, indicating repolarization issues, shows moderate importance in cluster 2 but is confounded by CVD in cluster 3. QRSE4, suggesting conduction alterations from fibrosis, is key in low-CVD cluster 1 but masked in CVD-heavy cluster 3.

Figure 2 Top 15 features for predicting diabetes mellitus in cluster 4. The feature importance plot from the gradient boosting machine model shows the relative contribution of each variable. Tfi: T-wave flatness index; PpeakN: Negative P-wave peak amplitude; Rpeak: R-wave amplitude; Pst: P-wave start time; SDNN: Standard deviation of normal-to-normal intervals; HFQRS: High-frequency QRS; QTc: Corrected QT; QRSfi: QRS complex morphology; SA: Amplitude of the S wave; QRSE3: To the ranges set by the frequency grid of 2-4-8-16-32 Hz; J80A: Amplitude at point J+80 milliseconds, μV.

The recurrence of features like Tfi and QTc confirms their link to DM pathophysiology (e.g., neuropathy, fibrosis). However, their diagnostic power is phenotype-dependent, rendering universal screening models ineffective. Stratification is essential - for example, Tfi + atrial features [P-wave amplitude (PpeakN), Pst] achieve optimal performance (AUC: 0.88) in cluster 4.

Tfi is a universal DM biomarker but is clinically reliable only when DM (not CVD) primarily drives ECG changes. Cluster 4 (high DM, moderate CVD) is ideal for ECG-based DM screening due to clear DM-specific feature expression. Unique features add critical insights: Atrial changes (PpeakN, Pst) in cluster 4 imply early DM impact, while QRS complex morphology dominance in cluster 2 may mimic DM effects.

DISCUSSION

The present study employed an innovative clustering approach to address a fundamental methodological challenge in detecting DM via single-lead ECG: The confounding influence of underlying CVD. Our initial analysis revealed that ECG alterations likely reflected broader CVD pathology rather than DM-specific signatures, as DM frequently presents as a comorbid condition rather than an isolated factor. This necessitated a paradigm shift beyond conventional DM vs non-DM comparisons. By clustering the entire cohort based on comprehensive clinical and demographic profiles, we identified four distinct phenotypic subgroups: Younger healthy individuals (cluster 1), older healthy individuals (cluster 2), and two groups with significant comorbidities but divergent profiles (clusters 3 and 4). This stratification proved crucial for disentangling the complex interplay between DM and CVD and for developing targeted diagnostic algorithms.

Cluster 4 emerged as the most informative subgroup for investigating DM-specific ECG signatures. While both clusters 3 and 4 comprised patients with substantial comorbidity burdens, cluster 4 was distinguished by the highest prevalence of DM (39.1%), a predominance of males (75.0%), a relatively preserved mean ejection fraction (58.7% ± 7.7%), and moderate levels of left ventricular diastolic dysfunction (52.2%) and valvular pathologies (e.g., mitral stenosis categories). Crucially, compared to cluster 3, cluster 4 exhibited less pronounced cardiac structural remodeling (lower mean end-diastolic volume: 101.0 ± 42.3 mL vs 113.1 ± 51.4 mL) and a lower prevalence of severe heart failure (chronic heart failure_2/3/4). This profile suggests cluster 4 represents patients where DM is a major contributing factor within a context of significant, but not maximally severe, CVD, potentially allowing DM-related electrophysiological changes to manifest more distinctly than in cluster 3, where advanced structural heart disease may dominate the ECG signal.

The cluster-stratified model validation further underscored the significance of phenotype-specific analysis and the relevance of cluster 4. While the model trained on cluster 1 (younger/healthy) achieved the highest AUC (0.947), its clinical utility for DM screening in broader, higher-risk populations is likely limited. Conversely, the model for cluster 3 performed poorly (AUC: 0.589), likely because severe coexisting CVD pathologies masked or overwhelmed any specific DM-related ECG features. The cluster 4 model, however, demonstrated a balanced and robust performance (AUC: 0.880, sensitivity: 0.750, specificity: 0.833, positive predictive value: 0.875, negative predictive value: 0.682), making it the most suitable candidate for real-world application in a comorbid population. Feature importance analysis within cluster 4 revealed that Tfi was the most influential parameter, followed by age, PpeakN, R-wave amplitude, and Pst. This contrasts sharply with other clusters; for instance, QRS complex morphology was dominant in cluster 2, while Tfi was also top in cluster 3 but with drastically different performance, reinforcing the context-dependence of ECG feature interpretation. The prominence of repolarization (Tfi) and atrial activity (PpeakN, Pst) features in cluster 4 aligns with known DM pathophysiology, including autonomic neuropathy and subtle conduction system alterations.

Our findings, which highlight the importance of repolarization (Tfi) and atrial conduction (PpeakN, Pst) features, resonate with - yet significantly extend - previous research on ECG alterations in DM. Traditional 12-lead ECG studies have long reported associations between DM and repolarization abnormalities, but with considerable inconsistency. For instance, while some studies in relatively healthy diabetic cohorts have noted flattened T-waves and altered repolarization heterogeneity[16,17], others in more advanced disease states have reported prolonged QRS duration and QTc intervals[18,19]. This inconsistency likely stems from the confounding effect of varying degrees of comorbid CVD, which our phenotypic stratification explicitly addresses. Studies that attempted to link QTc prolongation to DM often found it was more pronounced in patients with complications or severe hypoglycemia[20,21], suggesting it may be a marker of advanced disease rather than early detection. In contrast, our model identified Tfi - a more nuanced measure of T-wave morphology - as a top feature in a comorbid but structurally compensated cohort (cluster 4), suggesting it may capture subtler, earlier repolarization changes before they manifest as overt QTc prolongation. Similarly, the significance of atrial parameters like PpeakN and Pst in our model hints at early electropathology related to diabetic autonomic neuropathy impacting the atria, a dimension less frequently explored in conventional 12-lead analyses. Therefore, our work moves beyond these historical inconsistencies by demonstrating that when analyzed with ML and appropriate phenotypic context, single-lead ECG can distill a coherent DM-specific signature from previously deemed ‘non-specific’ findings.

Beyond the established parameters like QTc and T-wave morphology, our model identified several high-dimensional features, such as HFQRS and QRSE4, as important discriminators. While these lack direct traditional ECG correlates, their importance can be interpreted within the known pathophysiology of DM. HFQRS are known to reflect the integrity of ventricular depolarization; their reduction is associated with slowed conduction velocity and fragmentation of the electrical wavefront, often caused by myocardial fibrosis or ischemia. QRSE4 similarly captures alterations in the VAT sequence. In the context of DM, chronic hyperglycemia promotes the accumulation of advanced glycation end-products and reactive oxygen species, leading to diffuse interstitial fibrosis and electrical remodeling of the myocardium. This subclinical cardiomyopathy manifests as conduction heterogeneity, which ML models can detect as changes in the high-frequency energy and complexity of the QRS complex. Thus, the prominence of HFQRS and QRSE4 in our models provides a plausible, data-driven link to the fibrotic and conductive tissue changes that constitute the diabetic heart’s electrophysiological substrate, often preceding overt systolic dysfunction.

While the model for cluster 1 achieved the highest AUC (0.947), its exceptional performance must be interpreted within its clinical context. Cluster 1 comprises younger, predominantly healthy individuals with a low burden of CVD. In this low-prevalence, low-complexity population, the clinical utility of a DM screening tool is inherently limited, as the pre-test probability of disease is low and the incremental value over routine care is questionable. The model’s near-perfect performance in this cohort, while methodologically validating our approach, may also be less generalizable to real-world screening scenarios that typically target older, higher-risk populations. In contrast, we deliberately selected the cluster 4 model (AUC: 0.880) as optimal because it addresses a far more pressing clinical need: Accurate DM identification within a complex, comorbid patient profile where traditional risk factors are entangled and the pre-test probability is high. The robust performance of the cluster 4 model in this challenging context - characterized by a high prevalence of DM (39.1%), hypertension, and other structural heart diseases - demonstrates its potential for tangible clinical impact. It is precisely in these complex patients, often encountered in cardiology and primary care settings, that a rapid, non-invasive tool for detecting occult DM or pre-diabetes would be most valuable for guiding further diagnostic testing and intensifying preventive management.

These findings carry significant implications. Firstly, they strongly argue against a “one-size-fits-all” approach to ECG-based DM detection. Phenotypic stratification, as achieved through clustering, is essential to account for the modifying effects of age and comorbid CVD on the ECG. Secondly, cluster 4 appears to encapsulate a patient phenotype where DM exerts a detectable, modulatory influence on the ECG amidst prevalent CVD. The robust performance of the cluster 4-specific model suggests that single-lead ECG holds promise for DM identification within this specific, clinically relevant subgroup characterized by significant comorbidity burden where DM is a prominent factor. Future research should focus on validating this cluster 4 model in external cohorts with similar phenotypic profiles and exploring longitudinal outcomes. Refining the feature set, potentially incorporating additional photoplethysmography-derived parameters or serial measurements, may further enhance performance. Ultimately, this phenotype-targeted strategy offers a more realistic and potentially effective pathway towards implementing ECG-based DM screening tools in complex clinical settings.

The changes in the ECG in patients with type 2 DM on the 12-lead ECG considered not specific. Consistent research findings indicate that patients with DM without cardiovascular complications exhibit several electrocardiographic abnormalities, including tachycardia, shortened QRS and QT intervals, increased QT interval dispersion, reduced depolarization wave amplitudes, and accelerated ventricular myocardial activation time. Additionally, flattening of T wave, supported by diminished maximum and minimum values in repolarization body surface isopotential maps[22]. However, recent single center prospective study demonstrated that patients with type 2 DM experience prolongation in the QRS duration and QTc[23].

Notably, the ECG-only model’s performance is striking given the historical view that ECG changes in DM are “non-specific”. Prior studies using 12-lead ECGs reported inconsistent alterations (e.g., QRS shortening vs prolongation, variable QT dynamics), often attributed to advanced cardiovascular complications[24-27]. Our results challenge this paradigm, indicating that single-lead ECG biomarkers - when analyzed via ML - can detect subclinical autonomic dysfunction before overt symptoms arise. This aligns with recent evidence linking DM to prolonged VAT and repolarization heterogeneity (Tfi), driven by hyperglycemia-induced fibrosis and ion channel remodeling[28].

The ECG parameters predictive in our model reflect early cardiac autonomic neuropathy, a well-established complication of DM[15,29,30]. For instance, Tfi and reduced Tpeak correlate with sympathetic denervation, diminishing T-wave amplitude[31,32]. Similarly, increased QRSE4 signifies aberrant conduction velocity due to myocardial collagen deposition, while VAT prolongation aligns with delayed depolarization from hyperglycemia-induced sodium channel dysfunction. These findings support the hypothesis that cardiac electrophysiological changes precede structural damage, positioning ECG as a viable biomarker for early DM detection. Our results corroborate Kittnar’s observations of repolarization abnormalities in DM patients without CVD, while extending them to single-lead applications[1].

Feature repetition underscores shared DM pathophysiology, but diagnostic utility is context-dependent. Phenotypic clustering (e.g., prioritizing cluster 4) prevents misattributing CVD-driven ECG changes to DM, enabling accurate screening. Future work should validate findings externally and explore causal relationships through longitudinal studies.

The Qardio-Qvark^® device can be used as a screening tool for early detection of DM. The advantages of using of this tool include cost effective, easy to use and can be implemented in the outpatient department as well as in-patients department in addition to at home use. The patient can record the single lead ECG during one minute and get the results immediately and if some alterations detected in the obtained results that generated by the algorithms implemented in the device, further, the results can be sent to the responsible physician and consulted accordingly. Moreover, the device is currently taught to diagnose several conditions such as, heart failure in both types systolic and diastolic types, ischemic heart disease, arrythmias, and diastolic dysfunction. Therefore, it is more comfortable to do general population screening for multiple organs from a single minute recording a single lead ECG.

Several limitations of this study must be acknowledged. First, the single-center design at a tertiary university hospital in Moscow and the homogeneous Russian cohort limit the generalizability of our findings. The patient population, clinical practices, and genetic background may differ from other healthcare settings and ethnic groups, potentially affecting the performance of our phenotype-stratified model. External validation in multi-center, international, and more diverse populations is essential to confirm the robustness and transportability of our approach.

Second, the representativeness of our sample must be considered. As a hospital-based cohort, our participants likely have a higher burden of comorbidities and more advanced disease stages than the general population or a primary care setting. This spectrum bias means our model, particularly the high-performing cluster 4 algorithm, is optimized for a comorbid, high-prevalence population and may not perform as well in a healthier, community-based screening context.

Third, our methodology is inherently tied to the specific hardware and software of the Qardio-Qvark^® device. The 47 ECG parameters used for model training are derived from proprietary algorithms. This device-specificity means our model is not directly applicable to raw ECG signals from other single-lead or standard 12-lead ECG systems without significant recalibration and validation. Furthermore, while we excluded recordings with poor quality, the finger-based form factor remains susceptible to motion artifacts and noise, which could influence parameter extraction.

Fourth, the clustering was based on available clinical and demographic variables; unmeasured confounders (e.g., detailed socioeconomic status, specific dietary habits, or genetic predispositions) may influence phenotypic stratification and introduce residual confounding. Fifth, the cross-sectional design precludes assessment of temporal ECG changes or DM progression. Finally, while the overall sample size is substantial, the subsequent division into four clusters resulted in modest subgroup sizes (e.g., cluster 4, n = 92), which may reduce the statistical power for cluster-specific analyses and increase the risk of overfitting, despite our use of robust cross-validation techniques. Further investigations required to confirm the clinical validity of the built ML model and external validation is the cornerstone for future studies.

CONCLUSION

This study demonstrates that a phenotype-stratified approach is crucial for effective ECG-based DM detection, with the model for cluster 4 (high DM prevalence, significant comorbidities) achieving robust performance (AUC: 0.880). This confirms that single-lead ECG can accurately identify DM within a specific, clinically relevant high-risk phenotype where DM is a prominent driver of electrophysiological changes.

The clinical implementability of this strategy is a key advantage. By utilizing a low-cost, portable device that requires only a one-minute recording, our approach moves beyond proof-of-concept towards a practical screening solution. This makes it particularly suitable for resource-limited primary care settings, community health programs, and even remote patient monitoring. It can act as a high-throughput triage tool, enabling healthcare workers to identify at-risk individuals efficiently and direct confirmatory testing (like HbA1c or fasting plasma glucose) more strategically. By prioritizing a high-risk phenotype, this strategy maximizes the positive predictive value of screening, offering a scalable and cost-effective method to improve early DM detection in diverse healthcare environments. The need for future multi-center validation is important to confirm the clinical validity of our findings.