A Structured Critical Review of Machine Learning Approaches for ECG-Based Detection of Dysglycemia and Their Translational Readiness

Abstract

This structured critical review provides a comparative and analytically grounded overview of machine learning (ML) approaches for electrocardiography (ECG)-based detection of dysglycemia, with a specific focus on translational readiness for clinical screening. A structured literature search across PubMed, Scopus, Web of Science, and IEEE Xplore identified 183 records, of which 17 studies were included following predefined screening criteria and PRISMA-guided selection principles. The included studies demonstrate substantial heterogeneity in dataset size (ranging from <50 to >25,000 subjects), ECG acquisition modalities (single-lead, 12-lead, wearable), feature representations (raw signals, heart rate variability, engineered features), and ML strategies (classical algorithms, deep learning, and multimodal models). Reported model performance is generally high, with accuracy values frequently exceeding 0.85 and area under the curve (AUC) ranging from 0.78 to 0.99. Smaller experimental studies often report inflated performance (up to 96–99% accuracy), whereas large-scale population-based investigations demonstrate more moderate but clinically plausible results (AUC ≈ 0.80–0.85). External validation, a key requirement for clinical applicability, was performed in only a limited subset of studies (approximately 12%). From a physiological perspective, ML models exploit ECG alterations associated with dysglycemia, including reduced heart rate variability, QT interval prolongation, and changes in ventricular depolarization and repolarization dynamics. However, the relationship between metabolic dysfunction and ECG signals remains indirect. A key finding of this review is the mismatch between reported predictive performance and translational readiness. The majority of studies (≈65–70%) are classified as early-stage (Level 1–2 or 2–3), relying on small, single-center datasets and internal validation. Only a minority of studies achieve near-translational maturity (Level 4), characterized by large-scale datasets and external validation. ECG-based dysglycemia detection represents a promising non-invasive and scalable screening paradigm. However, its clinical translation is constrained by the lack of standardized ECG acquisition protocols, limited dataset diversity, insufficient external validation, and fragmented methodological approaches. Future research should prioritize large multi-center datasets, standardized feature extraction pipelines, hybrid interpretable models, and prospective validation to enable robust, generalizable, and clinically deployable screening systems.

1. Introduction

Type 2 diabetes has emerged as one of the most pressing global health challenges of the 21st century and remains a leading cause of mortality, disability, and reduced quality of life worldwide. Over recent decades, the burden of type 2 diabetes has increased at an unprecedented rate. Epidemiological estimates indicate that the global number of individuals affected by type 2 diabetes rose from 148.4 million (135.5–162.6 million) in 1990 to 437.9 million (402.0–477.0 million) in 2019 [1], representing nearly a threefold increase within a single generation. This rapid growth is largely attributed to a combination of demographic and lifestyle transitions, including population aging, urbanization, reduced physical activity, and the global rise in obesity. Importantly, the burden of diabetes is no longer confined to older populations [2].

Despite advances in pharmacological treatment and disease management, early detection of type 2 diabetes remains a critical unresolved problem. The disease is characterized by a prolonged asymptomatic phase, during which metabolic dysregulation gradually progresses without overt clinical signs. As a result, a substantial proportion of individuals remain undiagnosed until the development of complications affecting multiple organ systems, including neuropathy, nephropathy, retinopathy, and cardiovascular disease [3,4,5]. From a clinical perspective, this delay in diagnosis significantly limits the effectiveness of preventive strategies and increases both morbidity and healthcare costs. From a public health standpoint, it underscores the need for scalable, accessible, and cost-effective screening approaches capable of identifying individuals at risk at earlier stages of disease progression.

Current diagnostic standards are based on laboratory measurements of blood glucose and glycated hemoglobin (HbA1c), which serve as the clinical reference for diabetes diagnosis [6]. Additional tools, such as capillary glucometers and continuous glucose monitoring systems, are widely used for disease monitoring and management [7,8]. While these methods provide accurate and clinically validated measurements, their implementation in large-scale screening remains constrained by practical considerations. Blood sampling, the need for laboratory infrastructure, device costs, and issues related to patient compliance limit their feasibility, particularly in low-resource or geographically remote settings. These limitations have driven increasing interest in the development of non-invasive, easily deployable technologies that could enable population-level screening without the need for biochemical testing.

Among such approaches, the use of Electrocardiography has attracted growing attention [9]. Electrocardiography is a widely available, low-cost, and non-invasive technique routinely used in clinical practice and increasingly integrated into wearable devices. The rationale for its application in diabetes screening is based on the well-established impact of metabolic disorders on cardiac electrophysiology. Diabetes is known to influence cardiac function through multiple interconnected mechanisms, including autonomic neuropathy [10], metabolic disturbances [11], and structural alterations of the myocardium [12]. These processes lead to measurable changes in electrocardiographic parameters, such as heart rate variability and ventricular repolarization [13].

One of the earliest and most clinically relevant manifestations is cardiovascular autonomic neuropathy, which results from chronic hyperglycemia-induced damage to autonomic nerve fibers regulating cardiac function [14]. The extensive innervation of the heart involves a complex network of sympathetic and parasympathetic pathways, and disruption of this system leads to impaired regulation of cardiac output and electrophysiological stability [15]. A key measurable consequence of autonomic dysfunction is a reduction in heart rate variability (HRV), which has been consistently observed in individuals with diabetes and even in prediabetic states [16]. Importantly, HRV parameters show significant associations with glycemic markers, including fasting glucose and HbA1c, as well as with disease duration, suggesting a progressive deterioration of autonomic regulation over time. Structural and functional alterations in cardiac parasympathetic pathways, including postganglionic neurons within intracardiac ganglia, contribute to the withdrawal of parasympathetic tone and increased susceptibility to arrhythmias [17].

At the cellular level, metabolic disturbances associated with diabetes alter the function of cardiac ion channels. Chronic hyperglycemia, together with oxidative stress and inflammatory processes, modifies the activity and expression of sodium, potassium, and calcium channels [18,19]. These alterations disrupt the normal dynamics of the cardiac action potential, affecting both depolarization and repolarization processes. In particular, impaired potassium channel function is associated with delayed repolarization and prolongation of the QT interval, a well-recognized electrocardiographic marker linked to increased risk of arrhythmias and sudden cardiac death [20]. Clinical studies have demonstrated that QT interval duration correlates with HbA1c levels and disease duration, indicating that poor glycemic control may directly influence cardiac electrophysiological stability [21]. In addition, changes in QRS duration and QT dispersion have been reported, particularly in patients with longer disease duration, further suggesting increased heterogeneity in electrical conduction [22].

Beyond cellular mechanisms, diabetes also induces structural and functional alterations in the myocardium, often described as diabetic cardiomyopathy. This condition is characterized by myocardial fibrosis, hypertrophy, and impaired calcium handling, resulting from dysfunction of key regulatory proteins such as ryanodine receptors, sarcoplasmic reticulum calcium ATPase, and sodium–calcium exchangers [23]. These changes lead to impaired contractility and altered propagation of electrical signals within cardiac tissue. At the same time, chronic inflammation and oxidative stress contribute to electrical remodeling, while microvascular dysfunction impairs myocardial perfusion and may lead to subclinical ischemia [24]. The combined effect of these processes is the emergence of subtle, heterogeneous, and often non-specific alterations in ECG signals. To provide a structured overview of these interconnected mechanisms and their implications for ECG-based screening, a schematic representation is presented in Figure 1.

The convergence of these mechanisms has motivated the exploration of machine learning techniques applied to ECG data for the detection of diabetes and related metabolic abnormalities [25]. In principle, ECG-based screening offers several attractive features, including non-invasive acquisition, low operational cost, and compatibility with large-scale deployment through wearable technologies. Despite growing interest in ECG-based detection of dysglycemia, the existing literature remains fragmented and methodologically heterogeneous. Moreover, many published studies remain preliminary, exploratory, or insufficiently reported for rigorous methodological comparison. Previous reviews have summarized mechanistic links and ECG biomarkers in dysglycemia, including our prior work [26], which focused primarily on pathophysiological and electrophysiological aspects. However, these studies did not systematically evaluate machine learning methodologies, dataset characteristics, and translational readiness. Studies differ substantially in dataset scale, population characteristics, ECG acquisition protocols, feature representation, and validation strategies.

These pathophysiological mechanisms may contribute to ECG-derived feature alterations commonly analyzed in machine learning studies, including HRV metrics, QT-related parameters, and waveform morphology descriptors.

In this context, this review aims to provide a structured critical synthesis of machine learning approaches for ECG-based dysglycemia detection. Specifically, it categorizes existing studies by data sources, ECG modalities, feature representation, and modeling strategies, and evaluates reported performance in relation to dataset characteristics and validation design. The novelty of this work lies in its translational perspective. Beyond summarizing existing methods, this review analyzes how key methodological factors influence model robustness and proposes a conceptual translational readiness categorization intended to support comparative interpretation of methodological maturity and clinical applicability across current approaches. The proposed translational readiness categorization is intended as a conceptual analytical tool rather than as a formally validated clinical evaluation framework. This integrated analysis provides a clearer understanding of the capabilities and limitations of ECG-based screening and outlines directions for future research toward scalable and clinically relevant solutions.

2. Literature Search and Review Methodology

2.1. Search Strategy

To identify and critically evaluate studies on machine learning approaches for ECG-based detection of dysglycemia, a structured literature search was conducted across major biomedical and engineering databases, including PubMed, Scopus, Web of Science, and IEEE Xplore. The search strategy targeted studies at the intersection of electrocardiography, glycemic disorders, and artificial intelligence, using keywords related to ECG (“ECG”, “electrocardiogram”, “heart rate variability”), dysglycemia (“diabetes”, “prediabetes”, “hyperglycemia”, “dysglycemia”), and machine learning (“machine learning”, “deep learning”, “artificial intelligence”). These terms were combined using Boolean operators as follows: (“ECG” OR “electrocardiogram” OR “heart rate variability”) AND (“diabetes” OR “prediabetes” OR “hyperglycemia” OR “dysglycemia”) AND (“machine learning” OR “deep learning” OR “artificial intelligence”). In addition, several targeted search phrases were used to capture variations in terminology across studies, including “ECG-based diabetes detection”, “electrocardiogram diabetes prediction”, “ECG-based diabetes screening”, “ECG machine learning diabetes”, “non-invasive diabetes detection ECG”, “ECG signals diabetes classification”, and “type 2 diabetes detection deep learning ECG”. Search fields included title, abstract, and keywords where available. In PubMed, the search was performed in title/abstract fields; in Scopus, title–abstract–keywords were searched; in Web of Science, the Topic field was used; and in IEEE Xplore, metadata and abstract fields were searched. The database-specific search fields, Boolean query structures, and applied filters are summarized in

Table 1. Database-specific search strategy.

Database	Search Fields	Search Query	Filters/Limits
PubMed	Title/Abstract	(“ECG” OR “electrocardiogram” OR “heart rate variability”) AND (“diabetes” OR “prediabetes” OR “hyperglycemia” OR “dysglycemia”) AND (“machine learning” OR “deep learning” OR “artificial intelligence”)	English language; peer-reviewed articles; searched February 2025
Scopus	Title, Abstract, Keywords	TITLE-ABS-KEY (“ECG” OR “electrocardiogram” OR “heart rate variability”) AND TITLE-ABS-KEY (“diabetes” OR “prediabetes” OR “hyperglycemia” OR “dysglycemia”) AND TITLE-ABS-KEY (“machine learning” OR “deep learning” OR “artificial intelligence”)	English language; articles and conference papers; searched February 2025
Web of Science	Topic field	TS = (“ECG” OR “electrocardiogram” OR “heart rate variability”) AND TS = (“diabetes” OR “prediabetes” OR “hyperglycemia” OR “dysglycemia”) AND TS = (“machine learning” OR “deep learning” OR “artificial intelligence”)	English language; articles and proceedings papers; searched February 2025
IEEE Xplore	Metadata and Abstract	(“ECG” OR “electrocardiogram” OR “heart rate variability”) AND (“diabetes” OR “prediabetes” OR “hyperglycemia” OR “dysglycemia”) AND (“machine learning” OR “deep learning” OR “artificial intelligence”)	English language; journals and conference proceedings; searched February 2025

.

Additional targeted search phrases were also used to capture terminology variations across studies, including “ECG-based diabetes detection”, “electrocardiogram diabetes prediction”, “non-invasive diabetes detection ECG”, and related combinations. Reference lists of relevant studies were additionally screened to improve coverage. Google Scholar was additionally used for supplementary citation screening and identification of potentially relevant studies. Given the exploratory and interdisciplinary nature of this structured critical review, a formal review protocol was not registered. However, predefined search and screening principles were applied to ensure methodological consistency and transparency.

The search strategy was intentionally designed to prioritize studies explicitly combining ECG-based analysis, machine learning methodologies, and quantitative dysglycemia-related prediction outcomes, thereby excluding broader cardiovascular AI studies without direct glycemic relevance.

2.2. Eligibility Criteria

Study selection was guided by predefined inclusion and exclusion criteria. Studies were included if they utilized ECG signals or ECG-derived features, including heart rate variability, as input data and applied machine learning or deep learning methods for the detection, classification, or prediction of diabetes, prediabetes, or dysglycemia. Eligible studies were required to report quantitative performance metrics such as accuracy, area under the curve (AUC), sensitivity, or specificity. Studies were excluded if they did not involve ECG data, did not employ machine learning approaches, or focused exclusively on Type 1 diabetes, gestational diabetes, or disease management rather than detection. Additionally, studies lacking performance evaluation, as well as review articles, editorials, and non-peer-reviewed preprints, were excluded. Given the interdisciplinary and exploratory nature of ECG-based dysglycemia detection research, the included studies encompassed multiple related clinical endpoints (diabetes, prediabetes, hyperglycemia, and dysglycemia), ECG representations (raw ECG, HRV-derived features, ECG images, and multimodal approaches), and analytical objectives (classification, regression, biomarker estimation, and risk prediction). Rather than treating these studies as directly comparable, the review employed a structured thematic synthesis approach in which studies were analyzed within conceptual methodological categories. Comparative interpretation focused primarily on methodological trends, validation strategies, translational limitations, and dataset characteristics rather than direct numerical performance comparison across heterogeneous prediction tasks.

Given the methodological heterogeneity and exploratory nature of the field, the review intentionally focused on studies that provided sufficiently detailed ECG-based machine learning methodologies together with quantitative performance evaluation relevant to dysglycemia detection. This selective approach was intended to support structured comparative analysis rather than exhaustive bibliographic coverage of all diabetes-related ECG studies.

2.3. Study Selection and Data Extraction

All records identified through database searches were subjected to a structured multi-stage screening process. Initially, duplicate entries were identified and removed. The remaining records were screened based on titles and abstracts to exclude clearly irrelevant studies. Subsequently, full-text articles were assessed for eligibility according to the predefined inclusion and exclusion criteria. The study selection process was documented using a PRISMA-guided flow diagram [27], with the corresponding PRISMA checklist available in Supplementary File S1, and is illustrated in Figure 2. A total of 183 records were initially identified, of which 52 duplicates were removed, resulting in 131 records for screening. Following title and abstract screening, 78 records were excluded. The remaining 53 articles were assessed in full text, and 36 studies were further excluded for not meeting the eligibility criteria. The majority of excluded studies did not directly address ECG-based dysglycemia detection using machine learning methods or lacked sufficient methodological and performance reporting for structured comparative analysis. Ultimately, 17 studies were included in the final analysis. Full-text exclusion decisions were grouped into four predefined categories: absence of ECG or ECG-derived input data, absence of machine learning or deep learning methodology, lack of quantitative performance metrics, and outcomes unrelated to dysglycemia detection or prediction. The number of excluded studies in each category is reported in Figure 2 to improve transparency of the selection process.

Data extraction was performed using a predefined extraction template to ensure consistency across studies. For each included study, the following information was collected: publication year, dataset characteristics, ECG acquisition type (single-lead, multi-lead, or wearable), feature representation (including raw signals, heart rate variability, or engineered features), sample size, population characteristics, glycemic markers (HbA1c, glucose levels, or diagnostic labels), study design, machine learning model, performance metrics, and reported limitations. The extracted data were used to construct the comparative summary presented in

Table 2. Operational criteria for translational readiness classification.

Level	Operational Characteristics
Level 1	Proof-of-concept studies with very small cohorts, highly controlled conditions, and no clinical validation
Level 2	Early-stage ML development studies with internal validation only and limited dataset diversity
Level 3	Retrospective clinical validation studies using larger or clinically structured datasets
Level 4	Studies with external validation and/or large population-based cohorts approaching translational applicability
Level 5	Prospective real-world clinical deployment or implementation studies

and to support a structured analysis of methodological trends and model performance across the included studies. Screening and data extraction were conducted by the primary reviewer and subsequently checked for consistency and completeness by the co-authors.

2.4. Methodological Limitations and Potential Sources of Bias

The included studies were critically appraised with respect to methodological transparency, validation design, dataset composition, and potential sources of bias relevant to ML-based ECG analysis. The assessment focused on the transparency and completeness of study reporting, including ECG data acquisition protocols, dataset sources, study populations, and clearly defined inclusion and exclusion criteria.

Additional evaluation considered the clarity of the prediction task, the description of preprocessing procedures, and the extent to which input features and model variables were explicitly reported. Particular attention was given to class distribution reporting, given the potential impact of class imbalance on model performance. The completeness and consistency of reported performance metrics, including accuracy, sensitivity, specificity, and AUC, were also examined. Studies were assessed based on the robustness of validation strategies, including the use of internal validation, cross-validation, or external validation on independent datasets. Particular attention was given to whether validation was performed at the patient-level or using segment-level/beat-level expansion strategies, because this distinction substantially influences the risk of optimistic bias and data leakage. Studies relying on small sample sizes, lacking external validation, or providing insufficient methodological details were considered to have a higher risk of bias. Because of the substantial heterogeneity in study objectives, ECG representations, prediction targets, and analytical designs, a formal standardized bias assessment tool (PROBAST or QUADAS-2) was not applied. Instead, studies were comparatively appraised using a structured methodological evaluation focused on dataset characteristics, validation rigor, reporting transparency, population representativeness, and potential overfitting risks.

2.5. Data Synthesis

The synthesis of the included studies was performed using a structured approach to enable systematic comparison of methodologies and results. The studies were analyzed using a structured thematic synthesis approach designed to compare methodological trends, dataset characteristics, feature representations, validation strategies, and translational limitations. The synthesis was organized around several key analytical dimensions relevant to ECG-based dysglycemia detection, including general study characteristics, dataset composition and population profiles, ECG acquisition protocols and signal configurations, feature representation approaches (such as raw signals, heart rate variability, and engineered features), machine learning strategies (including conventional algorithms, deep learning models, and multimodal approaches), and reported performance metrics together with validation schemes.

Each study was critically evaluated in terms of methodological clarity, consistency of reporting, and potential sources of bias, as described in the previous section. Performance indicators, including accuracy, sensitivity, specificity, F1-score, and area under the curve (AUC), were descriptively extracted to support contextual methodological interpretation. Because of substantial heterogeneity in prediction tasks, dataset composition, class distributions, validation strategies, and outcome definitions, these metrics were not treated as directly comparable quantitative estimates across studies. Reported performance metrics should therefore be interpreted only within study-specific methodological contexts and should not be viewed as indicators of comparative model superiority across fundamentally heterogeneous analytical paradigms.

To improve analytical consistency, studies were interpreted within thematic methodological subgroups, including:

-

ECG-based classification studies;

-

Biomarker estimation approaches (HbA1c or glucose prediction);

-

HRV-focused physiological modeling;

-

Multimodal prediction frameworks;

-

Wearable or continuous monitoring approaches.

Direct quantitative comparison between these categories was considered methodologically inappropriate due to differences in prediction targets, dataset composition, signal representations, and evaluation protocols. The results were structured into thematic analytical categories, including data sources, ECG acquisition, feature representation, machine learning approaches, and model performance.

To support structured interpretation of methodological maturity and translational applicability, the included studies were additionally categorized using a conceptual translational readiness framework. The classification was based on predefined operational criteria including dataset scale, validation strategy, population representativeness, and degree of clinical applicability. The proposed categorization is intended as an analytical comparative tool rather than a formally validated clinical evaluation framework. Operational criteria for each maturity level are summarized in Table 2.

3. Results

3.1. General Characteristics and Aims

Despite their shared use of ECG-derived information, the included studies represent fundamentally different methodological and translational paradigms rather than a single unified analytical domain. For conceptual clarity, the reviewed literature was interpreted within five distinct categories:

• ECG-only screening models aimed at non-invasive dysglycemia detection;
• Biomarker estimation approaches focused on HbA1c or glucose prediction;
• HRV-centered physiological modeling studies;
• Multimodal prediction systems incorporating ECG together with glycemic or clinical variables;
• Experimental proof-of-concept and highly controlled exploratory studies.

These categories differ substantially with respect to clinical objectives, physiological assumptions, input data structure, validation design, intended clinical application, and translational readiness. The studies were not interpreted as directly competing approaches, and reported performance metrics were considered only within their respective methodological contexts.

The studies included in this review (n = 17) provide a comprehensive overview of current machine learning approaches for ECG-based detection of dysglycemia. As summarized in

Table 3. Comparative summary of included studies.

Study	Year	Signal	Dataset	ECG	N	Population	Marker	Design	Model	Performance	Limitations	Maturity
[28]	2021	ECG	Outpatient cohort (hospital-based)	12-lead, 500 Hz, 10 s; intervals (HR, PR, QRS, QT, QTc), axes (P, QRS, T)	4832	Non-DM, prediabetes, T2DM; mean duration ~4.7 y	HbA1c	Retrospective cohort (validated)	CNN-based DL (ResNet + SE + attention)	AUC 0.826; Sens 71.9%; Spec 77.7%	Moderate accuracy; reduced performance in severe DM; single-center	Level 3 (retrospective clinical validation)
[29]	2023	ECG	Ethnic cohort (Sindhi, India; high-risk families)	12-lead, 10 s, 1000 Hz	1262 (10,461 beats)	Mean age ~48 y; 61% female; high cardiometabolic burden	HbA1c, FPG, RBG	Observational; train/val/test split	XGBoost (best); compared with RF, MLP, LSTM, CNN, Transformer	Acc 96.8%; Prec 97.1%; Rec 96.2%; F1 96.6%	Selection bias; no external validation; beat-level analysis; limited generalizability	Level 2 (model development, internal validation)
[30]	2022	ECG	Private dataset (non-public)	Single-lead; 256 Hz; resting	86 (24,630 segments)	35 T2DM/51 healthy; age 20–70 y	Glucose (≥160 mg/dL)	Supervised classification	Decision Tree (DTC); compared with FT, MT, CT	Acc 86.9%; Sens 81.9%; Spec 90.6%; F1 82.8%	Small sample; no external validation; private dataset; limited generalizability	Level 2–3 (prototype; limited clinical validation)
[31]	2021	ECG	Private dataset (Taiwan; ECG + glucose)	Single-lead; 1000 Hz; 60 s	1119	Age 38–80 y; mixed glycemic status	Blood glucose (≥100 mg/dL)	Retrospective; binary classification; 80/20 sp lit + CV	Deep NN (10-layer); compared with LR, SVM	AUC 0.945; Sens 87.6%; Spec 85.0%	Private dataset; no external validation; sensitive to signal quality	Level 3 (advanced ML validation)
[32]	2020	ECG	Self-collected dataset	3-electrode setup (wrist + ankles)	24 (~1500 samples)	10 diabetic/14 healthy	Clinical status (no HbA1c/glucose)	Experimental; 5-fold CV	SVM (cubic); compared with DT, LDA, NB, KNN	Acc 96.8%	Very small sample; no objective biomarkers; no external validation; high overfitting risk	Level 1–2 (proof-of-concept)
[33]	2021	ECG	Hospital-based dataset (wearable ECG)	Single-lead; 60 s segments	370 (~317k segments)	T2DM only; mean age ~43.5 y	HbA1c	Retrospective; 5-fold CV	CNN-MFVW; compared with CNN, CNN-LSTM	Acc 90.2%; AUC 0.990; F1 0.901	No control group; small cohort; no external validation; sensitive to preprocessing	Level 2–3 (model development; limited clinical validation)
[34]	2021	ECG	Self-collected experimental dataset	Single-lead; 1000 Hz	21 (~22k segments)	Young adults; mixed glycemic status	Blood glucose (OGTT)	Prospective; 3-class classification	DBSCAN + CNN	Acc 81.7%; Sens 98.5%; Spec 76.8%	Very small sample; controlled setting; selection bias; no external validation	Level 2 (early-stage experimental study)
[35]	2025	ECG + clinical (multimodal)	Population-based cohort (Qatar Biobank)	12-lead (clinical)	2043 + 395 (test)	Middle Eastern; mean age ~46 y	HbA1c, FPG	Cross-sectional + longitudinal (5-year follow-up)	DNN (ECG-DiaNet; ECG + CRFs)	AUC 0.845 (multimodal); 0.822 (CRF); 0.675 (ECG)	No external validation; single-region cohort; small longitudinal test set	Level 3 (advanced clinical ML; longitudinal validation)
[36]	2025	HRV (ECG-derived)	Retrospective cohort (AFT lab, India)	Lead II; 1000 Hz; 5 min segments	519 (261 T2DM/258 controls)	Age 18–55 y; no major comorbidities	FBG, PPBG, HbA1c	Retrospective; binary classification; 80/20 split	CatBoost (best); compared with LR, KNN, RF, GBM	Acc 91.3%; AUC 0.91; Sens 90.6%; Spec 91.9%	No external validation; controlled setting; HRV-only features; limited generalizability	Level 2–3 (validated ML model)
[37]	2025	ECG (engineered features)	Population-based cohort (Japan; external validation)	12-lead; 10 s; 500 Hz	16,766 + 2456 (external)	General population; higher risk in older subjects	FPG, HbA1c	Retrospective; internal + external validation	LightGBM (best); compared with LR, RF, XGBoost, DNN	AUC 0.851 (internal); 0.785 (external)	Feature-based (no raw ECG DL); moderate specificity; class imbalance	Level 4 (advanced clinical ML with external validation)
[38]	2022	ECG + demographics (multimodal)	EHR cohort (NYU Langone)	12-lead; 10 s; 250–500 Hz	25,951 (test); large training cohort	Outpatients; new-onset diabetes subgroup	HbA1c ≥ 6.5%	Retrospective; prediction; external validation	DL (ResNet); ECG + demographics	AUC 0.80 (model); 0.68 (risk score)	Selection bias; multimodal dependence; no real-world validation; data not public	Level 4 (advanced clinical ML; near-translational)
[39]	2022	ECG (image-based)	Hospital cohort (China; 3 centers)	12-lead ECG images; 5 s	~2914	Middle-aged/elderly; high-risk	FPG, OGTT	Retrospective; binary classification; CV + test set	CNN (JGRNet); compared with AlexNet, GoogleNet, SVM	Acc 0.781; AUC 0.777	Image-based ECG (information loss); no external validation; moderate performance	Level 2–3 (early DL with internal validation)
[40]	2023	Multimodal (ECG + glucose + ACC + respiration)	DINAMO wearable dataset (free-living)	Wearable ECG; 250 Hz; continuous (~4 days)	29 (20 healthy/9 diabetic)	Mixed cohort; continuous monitoring	Continuous glucose	Experimental; supervised classification	XGBoost (best); compared with LR, DT, RF, SVM	Acc 98.2% (multimodal); ~87.5% (ECG only)	Very small sample; uses glucose input; no external validation; high overfitting risk	Level 1–2 (exploratory multimodal study)
[41]	2024	ECG (high-density)	Private dataset (self-collected)	HD-ECG (up to 98 leads)	50	Healthy volunteers	Not specified	Experimental; supervised classification	CNN (HD-MVCNN)	Acc 99.0%; F1 94.5%	No glycemic ground truth; unclear labels; small sample; unrealistic setup (98 leads); no validation	Level 1 (concept study)
[42]	2023	ECG	MIMIC-III (ICU subset)	Single-lead; 125 Hz; 1 s windows	50	ICU patients; median age 64 y	Blood glucose	Retrospective; personalized classification	One-class SVM	AUC 0.92 (beat); 0.97 (10 s)	ICU-only cohort; small sample; personalized model; no external validation	Level 3 (advanced ML validation)
[43]	2017	HRV (RR-interval)	Public dataset (PhysioNet)	RR intervals (QRS-based)	50 (33 normal/17 diabetic)	Not specified	Not reported	Supervised classification	SVM	Acc ~95%	Very small sample; no glycemic markers; unclear labels; no external validation	Level 2 (early-stage study)
[44]	2024	HRV (ECG-derived)	Hospital cohort (Korea; prospective)	Wearable ECG; 250 Hz	83 → 21 (final)	T2DM only; elderly (mean ~69 y)	Continuous glucose	Observational; temporal prediction	1D CNN (ResNet-like; HRV input)	Acc 90.5%; Sens 87.5%; Spec 92.7%	Very small final cohort; no control group; HRV-only; no external validation	Level 3 (clinical ML validation)

, the selected literature spans multiple years and reflects the progressive development of this research field, from early studies based on small cohorts and handcrafted features to more recent investigations utilizing large-scale datasets and deep learning architectures. The included studies demonstrate substantial variability in dataset size, population characteristics, ECG acquisition modalities, feature representation strategies, and analytical objectives. This heterogeneity is consistent with the interdisciplinary nature of the field, which integrates cardiology, metabolic research, signal processing, and artificial intelligence. The diversity of methodological approaches and study designs underscores the need for a structured synthesis, which is further developed in the subsequent subsections focusing on datasets, ECG configurations, feature extraction, machine learning models, and performance evaluation. The characteristics of the included studies are summarized in Table 3.

According to the operational translational readiness framework summarized in Table 2, the majority of included studies corresponded to early-stage methodological categories (Level 1–3), whereas only a limited subset of investigations demonstrated characteristics approaching translational applicability through large-scale cohorts or external validation.

As shown in Table 3, the reviewed studies can be stratified into distinct conceptual methodological categories according to their primary clinical endpoints, ECG representations, and analytical objectives. A subset of studies focuses on the estimation of glycemic biomarkers, particularly glycated hemoglobin (HbA1c), directly from ECG signals, demonstrating that ECG-derived representations can capture clinically relevant metabolic information and may serve as surrogate markers for disease progression [28,33,38]. Another subgroup of studies focuses on classification of individuals into glycemic categories, including normoglycemia, prediabetes, and type 2 diabetes, where a wide range of machine learning techniques—such as gradient boosting, convolutional neural networks, and hybrid models—have been applied with frequently reported high performance under study-specific conditions [29,31,34,39,41].

In parallel, several studies adopt a feature-engineering approach based on physiological signal analysis, employing techniques such as intrinsic time-scale decomposition, empirical mode decomposition, and entropy-based feature extraction, followed by classification using conventional machine learning algorithms [30,32,42]. Additional research directions include heart rate variability–based modeling, which emphasizes autonomic dysfunction as a key mechanism associated with dysglycemia [36,43], as well as multimodal frameworks integrating ECG signals with clinical risk factors to improve predictive performance [35,44]. These multimodal and exploratory approaches were analyzed separately from ECG-only screening studies because their prediction targets, input structures, and translational assumptions differ substantially from purely non-invasive ECG-based detection paradigms.

Overall, the reviewed literature demonstrates that ECG-based dysglycemia detection has been explored across multiple methodological paradigms, including biomarker estimation, classification, physiological feature analysis, and multimodal prediction. This diversity of approaches highlights both the versatility of ECG as a data source and the absence of a unified modeling framework. Because these categories address fundamentally different prediction tasks and data modalities, the review emphasizes qualitative methodological interpretation rather than direct numerical comparison of reported performance metrics.

3.2. Data Sources and Study Populations

The studies included in this review exhibit substantial heterogeneity in both data sources and study populations, reflecting the early-stage and interdisciplinary development of ECG-based dysglycemia detection. As summarized in Table 3 and illustrated in Figure 3, the reviewed datasets span a wide spectrum, ranging from small self-collected experimental cohorts and private institutional databases to hospital-based clinical repositories, electronic health record (EHR) systems, public databases, and population-scale biobanks. A considerable proportion of studies relied on self-collected or non-public datasets with limited accessibility, including small experimental cohorts comprising 21–50 participants [32,34,40,41,43]. In contrast, other studies utilized more structured clinical data sources, such as hospital outpatient cohorts [28,33], retrospective institutional datasets [31,36], and EHR-based cohorts [38]. Publicly available databases, including MIMIC-III and PhysioNet, were employed in only a limited number of studies [42,43], while large-scale population-based datasets—such as the Qatar Biobank and a Japanese health checkup cohort with external validation—were used in relatively few investigations [35,37]. As shown in Figure 3, this distribution highlights the predominance of small-scale and institution-specific datasets, with comparatively limited use of large, diverse, and externally validated data sources.

A pronounced imbalance is also evident in the distribution of sample sizes across studies. Many investigations were conducted on small cohorts, typically including fewer than 100 subjects, often under controlled or experimental conditions [30,32,34,40,41,43,44]. In contrast, only a subset of studies leveraged larger datasets, including hospital-based cohorts with several thousand ECG records [28,33,39], a biobank-based dataset with more than 2000 participants [35], a large population-based cohort comprising 16,766 records with an additional external validation cohort of 2456 individuals [37], and an EHR-based study involving tens of thousands of patients [38]. This disparity indicates that much of the existing literature remains exploratory in nature, with limited statistical power and increased susceptibility to overfitting, whereas only a minority of studies approach the scale required for robust clinical validation and real-world deployment.

The composition of study populations further underscores the heterogeneity of the current evidence base. Several studies included mixed cohorts consisting of healthy individuals, prediabetic subjects, and patients with type 2 diabetes [28,29,36], whereas others focused on more specific or restricted populations, such as high-risk ethnic groups [30], outpatient cohorts [28], ICU patients [42], elderly individuals with established diabetes [44], or healthy volunteers [41]. In some cases, datasets were enriched with diabetic patients or lacked a control group entirely [33,44], thereby limiting their applicability to screening scenarios. Moreover, certain studies excluded participants with comorbidities [36] or were conducted under tightly controlled laboratory conditions [34,40], which may reduce ecological validity. The geographic distribution of datasets was also uneven, encompassing studies from Asia, the Middle East, India, the United States, and public repositories, but with limited cross-regional validation [35,37,38].

From a translational perspective, a key limitation of the reviewed literature lies in the mismatch between study populations and intended clinical applications. Many studies were conducted in small, highly selective, or experimentally controlled cohorts rather than in representative community-based populations. Only a limited number of investigations utilized large outpatient, biobank, or population-based datasets that more closely reflect real-world screening conditions [35,37,38], and external validation was performed in only a small subset of studies [37,38]. Taken together, these findings suggest that while ECG-based dysglycemia detection is technically feasible, the representativeness, diversity, and scalability of available datasets remain critical barriers to clinical translation.

3.3. ECG Acquisition and Signal Configuration

The reviewed studies demonstrate substantial variability in ECG acquisition protocols and signal configurations, which represents a key methodological factor influencing both feature representation and model performance. As outlined in Table 3, ECG data were acquired using a wide range of configurations, including standard 12-lead clinical ECG systems, single-lead recordings, high-density multi-lead systems, and wearable devices operating under both controlled and free-living conditions (Figure 4). This diversity reflects the flexibility of ECG as a sensing modality, but also highlights the absence of standardized acquisition protocols for dysglycemia detection, which complicates cross-study comparison and limits reproducibility.

A considerable proportion of studies relied on standard clinical 12-lead ECG systems, typically recorded over short durations of approximately 10 s under controlled conditions [28,29,37,38]. These datasets are generally associated with hospital-based or population-level cohorts and provide comprehensive spatial information on cardiac electrophysiology. In contrast, several studies employed single-lead ECG configurations, including both clinical and wearable setups, often with longer recording durations ranging from tens of seconds to several minutes [30,31,33,36]. While single-lead ECG enables simpler acquisition and improved scalability, it inherently reduces spatial information and may limit the detection of subtle electrophysiological changes associated with dysglycemia.

Wearable ECG devices represent an important emerging direction, particularly for continuous monitoring and real-world data collection. Several studies utilized wearable sensors under free-living or semi-controlled conditions, enabling long-term signal acquisition and temporal analysis of glycemic states [34,40,44]. These approaches are particularly relevant for large-scale screening applications, as they offer non-invasive and scalable data collection. However, wearable ECG recordings are more susceptible to motion artifacts, signal noise, and variability in acquisition conditions, which introduces additional challenges for preprocessing and model robustness.

In addition to conventional and wearable ECG configurations, a small number of studies explored alternative signal representations, including high-density ECG systems with up to 98 channels [41], ECG images instead of raw signals [39], and heart rate variability–based representations derived from RR intervals [36,43]. While these approaches provide additional perspectives on cardiac electrophysiology, they also introduce methodological inconsistencies and may limit comparability across studies. In particular, high-density ECG systems and image-based representations are less practical for real-world deployment, whereas HRV-based approaches depend heavily on preprocessing quality and may omit relevant waveform information.

Another critical source of variability lies in signal duration and segmentation strategies. Some studies analyzed short, fixed-length ECG recordings (10 s segments), while others employed longer recordings segmented into windows or heartbeat-level samples for model training [29,33,42]. In certain cases, beat-level segmentation was used to increase dataset size, although this may introduce data leakage or artificially inflate performance metrics if not properly controlled. Conversely, longer recordings enable more robust estimation of temporal features, such as heart rate variability, but may increase computational complexity and sensitivity to noise.

The reviewed literature indicates that ECG acquisition and signal configuration remain highly heterogeneous, with no consensus on optimal recording protocols for dysglycemia detection. This variability directly influences downstream feature extraction strategies and model design, as further discussed in the following subsection. Importantly, the lack of standardized ECG acquisition protocols represents a key barrier to reproducibility, comparability, and clinical translation of ECG-based dysglycemia detection systems.

3.4. Feature Representation and ECG-Derived Biomarkers

The representation of ECG signals and the selection of informative features constitute a central component of machine learning frameworks for dysglycemia detection. As summarized in Table 3, the reviewed studies employ a wide spectrum of feature extraction strategies, ranging from raw ECG signals and deep learning-based representations to engineered features derived from waveform morphology and heart rate variability (HRV). This methodological diversity reflects both the complexity of the underlying physiological mechanisms and the absence of a standardized feature representation framework for ECG-based dysglycemia detection.

A substantial group of studies relies on raw ECG signals as direct input to deep learning models, enabling automatic feature extraction without explicit signal engineering. Convolutional neural networks and related architectures were used to learn hierarchical representations of ECG waveforms, capturing complex temporal and morphological patterns associated with glycemic status [28,31,33]. These approaches are particularly advantageous in large-scale datasets, where sufficient data allow models to learn subtle, non-linear relationships between ECG signals and metabolic abnormalities. For example, deep learning models have been shown to estimate HbA1c levels and detect hyperglycemia directly from ECG signals, suggesting that ECG may encode latent biomarkers of glycemic regulation. However, the interpretability of such models remains limited, and their performance is highly dependent on dataset size and quality.

In contrast, several studies employ engineered features derived from ECG waveform characteristics, including classical electrophysiological parameters such as heart rate, PR interval, QRS duration, QT interval, QTc, and P- and T-wave morphology [28,29]. These features are physiologically interpretable and are directly linked to known mechanisms of diabetes-related cardiac dysfunction, including autonomic neuropathy, ion channel alterations, and myocardial remodeling. Feature-based approaches often incorporate preprocessing steps such as filtering, normalization, and fiducial point detection, followed by feature selection techniques to identify the most discriminative parameters [30,31]. While these methods provide greater transparency and clinical interpretability, they may fail to capture complex interactions present in raw signals.

Heart rate variability represents a distinct and widely used category of ECG-derived features, reflecting autonomic nervous system function. Multiple studies utilize time-domain, frequency-domain, and nonlinear HRV metrics, including SDNN, RMSSD, LF/HF ratio, and Poincaré plot parameters, to characterize alterations in cardiac autonomic regulation associated with dysglycemia [36,43]. These features are strongly grounded in physiological mechanisms, as reduced HRV is a well-established marker of autonomic dysfunction in diabetes. HRV-based models often achieve competitive performance and offer a simplified and computationally efficient representation of ECG data. However, they depend heavily on signal quality, accurate R-peak detection, and sufficiently long recording durations, and may omit important waveform-level information.

In addition to these primary approaches, several studies explore hybrid and alternative feature representations. Multimodal models combine ECG-derived features with clinical variables such as age, sex, and biochemical markers, demonstrating improved predictive performance compared to unimodal approaches [35]. Other studies employ feature-based machine learning pipelines using entropy measures, signal decomposition techniques (intrinsic time-scale decomposition or empirical mode decomposition), and statistical descriptors of ECG signals [30,32]. Furthermore, some approaches utilize ECG images or transformed representations rather than raw time-series data, although such methods may lead to information loss and reduced physiological interpretability [39].

Despite the diversity of feature extraction strategies, a common limitation across studies is the lack of standardization in feature definition, preprocessing pipelines, and evaluation protocols. Different studies use varying combinations of features, segmentation strategies, and normalization techniques, making direct comparison challenging. Moreover, the absence of consensus on which ECG-derived features are most relevant for dysglycemia detection reflects the complex and indirect relationship between metabolic disorders and cardiac electrophysiology. As discussed in Section 1, ECG alterations in diabetes are often subtle, non-specific, and influenced by multiple confounding factors, including comorbidities and inter-individual variability. An important limitation across the reviewed literature is the potential influence of confounding physiological and clinical factors on ECG-derived signatures associated with dysglycemia. Many electrophysiological alterations linked to diabetes, including reduced heart rate variability, QT interval prolongation, and conduction abnormalities, are not specific to glycemic dysfunction and may also reflect aging, hypertension, obesity, cardiovascular disease, autonomic imbalance, medication effects, or other systemic conditions. Consequently, the predictive information captured by machine learning models may partially reflect broader cardiometabolic risk profiles rather than glycemic status alone. Only a limited subset of studies explicitly controlled for potential confounding variables or evaluated model robustness across clinically heterogeneous populations.

Overall, the current evidence suggests that both deep learning–based representations and engineered physiological features can capture relevant information for dysglycemia detection, but each approach has inherent limitations. Raw signal–based methods offer higher flexibility and potential performance, whereas feature-based approaches provide greater interpretability and physiological grounding. HRV-based representations offer a simplified and clinically meaningful alternative but may sacrifice signal richness. The lack of unified feature representation strategies remains a major barrier to reproducibility and clinical translation, highlighting the need for standardized feature extraction frameworks and multimodal approaches in future research.

3.5. Machine Learning Models

The methodological landscape of machine learning approaches applied to ECG-based dysglycemia detection is characterized by a coexistence of classical algorithms and deep learning architectures, with no clear consensus regarding the optimal modeling paradigm. The choice of model is closely intertwined with data characteristics and feature representation, as discussed in Section 3.2, Section 3.3 and Section 3.4, and therefore reflects both the scale of available datasets and the underlying assumptions about signal informativeness.

A substantial portion of the reviewed studies employs conventional machine learning algorithms operating on engineered feature sets. These include decision trees, random forests, support vector machines, k-nearest neighbors, and gradient boosting methods such as XGBoost and LightGBM [29,30,36,37,42,43]. In many cases, these approaches are coupled with carefully selected physiological features derived from ECG waveforms or heart rate variability, allowing for relatively interpretable models that align with known mechanisms of diabetic cardiac dysfunction. Notably, gradient boosting methods frequently demonstrate strong performance in tabular settings, particularly when combining ECG-derived features with demographic or clinical variables [29,37]. However, their performance is inherently constrained by the quality and completeness of feature engineering, and their ability to capture complex temporal dependencies in raw ECG signals remains limited.

In parallel, deep learning models have been increasingly adopted, particularly in studies utilizing raw ECG signals or large-scale datasets. Convolutional neural networks (CNNs), including residual architectures, are the most commonly used models, enabling automated extraction of hierarchical features directly from waveform data [28,31,33,38,44]. These models are well-suited to capturing subtle and distributed patterns in ECG signals that may not be accessible through manual feature engineering. In some cases, deep neural networks have demonstrated the ability to infer glycemic status or estimate HbA1c levels from ECG data alone, suggesting that latent representations of metabolic state may be encoded in cardiac electrical activity. Nevertheless, such models often require large training datasets, and their performance may degrade significantly when applied to smaller or heterogeneous cohorts.

A smaller subset of studies explores hybrid and alternative modeling strategies, including multimodal architectures that integrate ECG features with clinical or demographic data [35], as well as specialized approaches such as one-class classification for personalized modeling [42] or clustering-assisted pipelines for preprocessing and feature selection [34]. These approaches reflect attempts to address specific limitations of conventional supervised learning, such as class imbalance, limited data availability, or inter-individual variability. While promising, such methods remain relatively underexplored and are often evaluated in small or highly specific cohorts, limiting the generalizability of their findings.

Despite the diversity of modeling techniques, several recurring methodological issues can be identified across the literature. First, the majority of studies rely on retrospective datasets and internal validation strategies, with only a limited number incorporating external validation on independent cohorts [37,38]. Second, class imbalance and selection bias are frequently insufficiently addressed, particularly in studies involving enriched or highly selective populations. Third, the risk of overfitting remains substantial, especially in studies with small sample sizes and high-dimensional feature spaces. Finally, the lack of standardized evaluation protocols and reporting practices complicates the comparison of model performance across studies.

Taken together, the current evidence suggests that both classical machine learning methods and deep learning architectures are capable of achieving high predictive performance under specific conditions. However, these results are highly dependent on dataset characteristics, feature representation, and validation strategy. In particular, models trained on small, homogeneous, or institution-specific datasets may not generalize to broader populations. Therefore, the primary challenge is not the absence of effective algorithms, but rather the development of robust, generalizable models supported by large, diverse, and externally validated datasets. Addressing these limitations will be essential for translating ECG-based machine learning models from proof-of-concept studies to clinically applicable screening tools. The choice of machine learning model appears to be secondary to dataset characteristics and feature representation, as model performance is primarily driven by data quality, scale, and validation strategy rather than algorithmic complexity.

3.6. Model Performance and Validation

The reported performance of machine learning models for ECG-based dysglycemia detection is generally high across the reviewed studies; however, these results should be interpreted with caution in light of substantial heterogeneity in evaluation protocols, dataset characteristics, and validation strategies. As summarized in Table 3, most studies report conventional performance metrics such as accuracy, sensitivity, specificity, and AUC, although interpretation of these values requires caution because evaluation protocols, prediction targets, and dataset structures differ substantially across studies.

A key observation emerging from the comparative analysis is the strong dependence of reported performance on dataset size and structure. As illustrated in Figure 5, studies based on small cohorts tend to report substantially higher performance metrics, often exceeding 0.90 or even 0.95, whereas studies utilizing larger and more heterogeneous datasets typically demonstrate more moderate results. This pattern suggests that model performance is not solely determined by algorithmic sophistication, but is strongly influenced by dataset scale, variability, and representativeness. In particular, small and homogeneous datasets may lead to overly optimistic estimates due to overfitting, limited variability, and reduced complexity of classification tasks.

Direct comparison of model performance across studies is further complicated by differences in problem formulation and outcome definitions. Some studies address binary classification tasks (diabetic vs. non-diabetic), while others consider multi-class classification or regression-based estimation of glycemic markers such as HbA1c. In addition, diagnostic thresholds vary across studies, including different cut-offs for fasting glucose or HbA1c levels, resulting in inconsistencies in labeling and evaluation. Consequently, similar numerical values of accuracy or AUC may correspond to fundamentally different prediction tasks and should not be interpreted as directly comparable measures of model effectiveness.

Another critical limitation lies in the design of validation strategies. The majority of studies rely on internal validation approaches, such as train–test splits or cross-validation applied to retrospective datasets. Although these methods are appropriate for initial model development, they are prone to optimistic bias, particularly when patient-level independence is not strictly enforced. In studies employing segmentation of ECG signals into multiple samples, the risk of data leakage increases if segments from the same individual appear in both training and testing sets. An important methodological distinction identified across the reviewed literature concerns the unit of validation. Several studies evaluated model performance at the beat-level or segment-level rather than at the patient-level, often substantially increasing the number of training and testing samples derived from a relatively small number of individuals. While such approaches may improve apparent statistical performance, they can introduce substantial risks of data leakage and optimistic bias if recordings from the same individual are represented in both training and testing subsets. Consequently, performance metrics derived from segment-level validation should not be interpreted as directly equivalent to patient-level clinical screening performance. Studies based on patient-level external validation generally reported more moderate but clinically plausible performance estimates, whereas studies relying on segment-level expansion or highly controlled cohorts frequently reported near-perfect classification metrics.

Dataset composition and population characteristics further affect reported performance. Several studies are conducted on enriched or highly selective populations, including high-risk groups, ICU patients, or cohorts without significant comorbidities. Several included studies should be interpreted primarily as exploratory or proof-of-concept investigations rather than clinically deployable screening models. In particular, some studies lacked clearly defined glycemic ground truth labels, included only diabetic cohorts without appropriate control populations, or incorporated glucose-related measurements directly into multimodal model inputs. While such studies provide insight into potential methodological directions, they are not directly representative of purely non-invasive ECG-based dysglycemia screening and therefore have limited translational applicability. In such cases, classification tasks may be artificially simplified, leading to inflated performance metrics that do not reflect real-world screening conditions. Conversely, studies based on large, heterogeneous populations tend to report lower but more realistic performance values, reinforcing the importance of dataset diversity for robust model evaluation.

Taken together, the evidence indicates that high model performance is achievable under controlled or small-scale conditions, but may not generalize to broader clinical populations. As highlighted in Figure 5, the apparent trade-off between dataset size and reported performance underscores the need for careful interpretation of published results. The lack of systematic confounder analysis across the reviewed studies further limits the interpretability of ECG-derived predictive signatures, as many electrophysiological features associated with dysglycemia overlap substantially with broader cardiovascular and metabolic risk factors. Consequently, reported model performance may partially reflect indirect associations with age, cardiovascular disease, autonomic dysfunction, obesity, medication effects, or other systemic conditions rather than glycemic status alone. Accordingly, high predictive performance in some studies should not necessarily be interpreted as evidence of dysglycemia-specific electrophysiological biomarkers, but may instead reflect indirect associations with broader cardiometabolic abnormalities. From a practical perspective, ECG-based dysglycemia detection is unlikely to replace established biochemical diagnostic methods such as fasting glucose or HbA1c testing. Instead, its most realistic clinical role may involve opportunistic pre-screening, population-level risk stratification, or continuous passive monitoring integrated into existing ECG workflows and wearable ecosystems. Potential use cases include preliminary screening during routine ECG examinations, wearable-based risk monitoring, remote health assessment in low-resource settings, and identification of individuals who may benefit from confirmatory biochemical testing.

In screening-oriented applications, sensitivity may be clinically more important than maximal specificity, since the primary objective would be early identification of potentially undiagnosed individuals rather than definitive diagnosis. Consequently, moderate specificity may be acceptable if balanced by sufficiently high sensitivity and low-cost confirmatory testing. However, excessively high false-positive rates could reduce clinical utility and increase unnecessary downstream diagnostic procedures.

For practical implementation, clinically acceptable model performance would likely require robust patient-level validation, stable performance across heterogeneous populations, and reproducible sensitivity–specificity trade-offs under real-world screening conditions. In addition to analytical performance, clinical deployment would require prospective multi-center validation, standardized ECG acquisition protocols, reproducible preprocessing pipelines, regulatory approval pathways, and demonstration of generalizability across diverse demographic and clinical populations. From a regulatory perspective, ECG-based machine learning systems intended for screening applications would likely require evaluation as software-as-a-medical-device (SaMD) platforms, including evidence of clinical safety, robustness, and post-deployment monitoring. At present, most reviewed studies remain at the retrospective proof-of-concept stage and do not yet provide sufficient evidence for immediate clinical implementation.

3.7. Model Maturity and Translational Readiness

The current body of evidence reveals a pronounced imbalance between reported model performance and the actual level of methodological maturity required for clinical translation. Despite consistently high predictive metrics across studies, the majority of approaches remain at early stages of development. As summarized in Table 3 and illustrated in Figure 6, most studies fall within Level 1–2 and Level 2–3 maturity, corresponding to proof-of-concept investigations and initial model development. These studies are typically based on small or highly controlled datasets and rely predominantly on internal validation strategies, which limits the reliability of their reported performance in real-world settings [32,34,40,41,43].

A subset of studies demonstrates a higher level of methodological rigor and can be classified within Level 3 maturity. These investigations generally utilize retrospective clinical datasets and apply more structured validation strategies, including cross-validation and independent test sets [28,31,33,36]. However, this group remains constrained by single-center data sources and limited population diversity. The absence of external validation in most of these studies raises concerns regarding their robustness and reproducibility across different clinical environments.

Only a small number of studies approach translational readiness (Level 4). These works are characterized by the use of large-scale datasets, such as population-based cohorts or electronic health record systems, and by the inclusion of external validation cohorts [35,37,38]. Such designs provide more realistic estimates of model performance and represent an important step toward clinical implementation. Nevertheless, even these studies remain predominantly retrospective and are often limited to specific geographic or institutional contexts, which restricts their broader applicability.

Taken together, the distribution of maturity levels presented in Figure 6 highlights a fundamental limitation of the field. High predictive performance is frequently achieved in low-maturity settings, whereas studies with more rigorous design and validation tend to report more moderate but clinically plausible results. This indicates that model maturity—defined by dataset scale, population representativeness, and validation strategy—provides a more meaningful indicator of translational potential than performance metrics alone. Advancing ECG-based dysglycemia detection will therefore require a shift toward externally validated, population-level, and prospectively evaluated models capable of integration into real-world clinical workflows.

4. Discussion

The present review indicates that ECG-based dysglycemia detection is a technically promising yet methodologically immature field. While many studies report high predictive performance, these results are largely driven by dataset characteristics, including limited sample sizes, controlled conditions, and population bias, rather than intrinsic model capability. At the same time, substantial heterogeneity persists in ECG acquisition protocols, feature representation strategies, and validation approaches, preventing meaningful comparison across studies. Importantly, most models remain at early or intermediate stages of maturity, with minimal external validation and limited alignment with real-world screening scenarios. Collectively, these findings suggest that the primary challenge is not the absence of informative ECG biomarkers, but the lack of standardized data acquisition, robust validation frameworks, and scalable systems required for clinical translation (

Table 4. Synthesis of key findings, limitations, and required improvements for clinical translation.

Domain	Key Finding	Strengths	Limitations	Required Improvement
Data characteristics	ECG-based dysglycemia detection is feasible across multiple datasets	Large-scale studies demonstrate predictive potential	Most studies rely on small, single-center datasets; limited diversity	Large, multi-center, population-level datasets
ECG acquisition	Signal characteristics strongly influence model performance	Multilead ECG provides richer physiological information	Heterogeneous acquisition protocols; frequent use of single-lead ECG	Standardized, wearable multilead ECG systems
Feature representation	Both engineered and deep features capture relevant information	HRV and repolarization features show physiological relevance	Lack of standardization; inconsistent preprocessing	Hybrid feature frameworks with standardized pipelines
Machine learning models	ML and DL models achieve high performance under controlled conditions	CNN, boosting models show strong results	Performance depends on dataset rather than model; limited interpretability	Robust, interpretable models validated across datasets
Model performance	High reported accuracy in many studies	Strong results in experimental settings	Overfitting, optimistic bias, poor comparability	Standardized evaluation metrics and protocols
Validation strategy	Validation is a key bottleneck	Some studies include external validation	Most rely on internal validation; data leakage risk	External and prospective validation
Model maturity	Majority of studies at early/intermediate levels	Emerging Level 3–4 studies	Limited translational readiness	Maturity-driven development frameworks
Clinical applicability	ECG has potential for non-invasive screening	Scalable and low-cost modality	No real-world deployment; lack of screening studies	Integration into clinical workflows and screening programs
System integration	End-to-end systems are required	Advances in wearable ECG devices	Fragmented pipelines; lack of standardization	Integrated acquisition–ML–validation systems

).

4.1. Requirements for Clinical Translation

The translation of ECG-based dysglycemia detection from experimental studies to clinically applicable screening systems requires the fulfillment of several methodological and technological criteria. Based on the analysis of the reviewed literature, these requirements can be structured into five key domains: data, acquisition, feature representation, modeling, and validation [45,46,47].

The availability of large-scale and representative datasets is a fundamental requirement. Models should be trained and evaluated on heterogeneous populations that reflect real-world screening conditions, including variability in age, sex, ethnicity, and comorbidities. The limitations of small and single-center datasets, which often lead to optimistic performance estimates, have been widely documented in machine learning–based medical studies [48]. In addition, standardized outcome definitions based on clinically validated biomarkers, such as HbA1c or oral glucose tolerance test (OGTT), are necessary to ensure consistency across datasets and enable meaningful comparison between studies [49].

ECG acquisition must be standardized, reproducible, and scalable. As discussed in Section 3.3, substantial variability exists in signal configurations, ranging from single-lead recordings to multilead clinical ECG systems. Previous studies have shown that simplified or non-standardized ECG acquisition may lead to information loss and reduced diagnostic reliability [50]. From a translational perspective, the use of multilead ECG systems with controlled signal quality, appropriate filtering, and consistent acquisition protocols is essential. At the same time, these systems must be designed for accessibility and deployment outside specialized clinical environments, which aligns with recent developments in wearable and portable ECG technologies [51,52].

Feature representation should balance physiological interpretability and robustness. While deep learning approaches enable automatic extraction of complex signal patterns, their reliance on large datasets and limited interpretability remain important challenges [53,54]. Conversely, engineered ECG features, including heart rate variability metrics and repolarization-related parameters, are grounded in well-established physiological mechanisms but may be sensitive to signal quality and preprocessing variability [55]. Hybrid approaches that combine data-driven and physiologically informed features have been increasingly proposed as a way to improve both performance and interpretability [56].

Rigorous validation strategies are required to ensure model generalizability. Internal validation methods, such as cross-validation, are insufficient to establish clinical applicability, particularly in the presence of dataset bias and potential data leakage [57,58]. External validation using independent datasets, preferably from different institutions or geographic regions, is widely recognized as a critical step in the development of reliable AI-based diagnostic systems [59]. Furthermore, prospective validation and real-world evaluation are necessary to assess performance under practical screening conditions and to account for variability in signal acquisition and patient behavior [60,61].

Clinical translation requires integration into scalable and user-friendly screening workflows. This includes not only model performance, but also system-level considerations such as ease of use, automation of signal processing, and compatibility with digital health infrastructures. The importance of end-to-end system design, encompassing data acquisition, processing, and interpretation, has been emphasized in recent studies on AI-based medical technologies [62]. In this context, the development of accessible ECG acquisition platforms combined with standardized analytical pipelines represents a key enabler for large-scale, non-invasive screening.

Taken together, these requirements define a structured pathway for advancing ECG-based dysglycemia detection toward clinical implementation. Addressing these conditions will be essential to bridge the gap between promising experimental results and reliable, scalable screening systems suitable for real-world use.

4.2. Toward Practical ECG-Based Screening Systems

The analysis presented in this review indicates that the primary barrier to the clinical adoption of ECG-based dysglycemia detection is not the absence of predictive signal, but the lack of an integrated and scalable framework that connects signal acquisition, feature representation, model development, and validation within a unified pipeline. Addressing this gap requires a transition from isolated, performance-driven studies toward system-oriented approaches that explicitly account for data quality, reproducibility, and deployment constraints. Such a transition is consistent with recent trends in digital health, where end-to-end system design has been recognized as a critical factor for successful clinical translation [63].

A central element of this transition is the standardization of ECG acquisition. As demonstrated in Section 3.3, many existing studies rely on heterogeneous signal configurations, including single-lead recordings or retrospective datasets acquired under uncontrolled conditions. From a screening perspective, this variability limits both reproducibility and physiological interpretability. A practical solution is the adoption of accessible wearable multilead ECG systems capable of providing clinically relevant signal fidelity while remaining suitable for large-scale deployment. Recent developments in wearable multilead ECG platforms, including systems capable of synchronized 12-lead acquisition, demonstrate the feasibility of combining clinical-grade signal quality with scalable deployment [64]. In particular, wearable 12-lead configurations with integrated filtering, signal quality control, and synchronized acquisition offer a promising compromise between clinical-grade diagnostics and scalability [65]. Such systems enable consistent extraction of multilead features, including repolarization indices, which may be sensitive to metabolic disturbances.

Recent developments in wearable multichannel ECG platforms further support this direction by demonstrating the feasibility of combining compact hardware design with structured signal processing and data management frameworks. Importantly, these systems should be viewed not as standalone diagnostic tools, but as standardized data acquisition platforms that facilitate the generation of high-quality datasets for subsequent analysis. This perspective aligns with the requirements outlined in Section 4.1, where the availability of reliable and reproducible input data was identified as a prerequisite for meaningful model development and validation.

Beyond acquisition, the integration of feature extraction and machine learning into a coherent analytical pipeline is essential. Rather than treating these components independently, future systems should ensure that feature representation is aligned with both signal characteristics and clinical objectives. Hybrid approaches that combine physiologically interpretable features with data-driven representations may offer a viable pathway toward robust and explainable models [66]. At the same time, model development must be coupled with rigorous validation strategies, including external and prospective evaluation, to ensure generalizability across populations and settings [67].

The overall pathway from ECG signal acquisition to clinically applicable screening is illustrated in Figure 7. This framework highlights the interdependence of key components, including standardized acquisition, feature consistency, model robustness, and validation rigor, as well as the methodological bottlenecks identified in the current literature. Notably, the incorporation of accessible wearable multilead ECG systems at the initial stage of the pipeline represents a critical enabler for overcoming existing limitations and supporting large-scale screening applications.

From a translational perspective, the development of practical ECG-based screening systems should prioritize accessibility, scalability, and integration into existing healthcare infrastructures. This includes the use of low-cost wearable devices, automated processing pipelines, and compatibility with digital health platforms. In addition, future research should focus on prospective validation studies and real-world deployment scenarios, particularly in resource-limited settings where non-invasive and scalable screening tools are most needed [68]. A proposed five-level conceptual translational readiness categorization for translational evaluation of ECG-based dysglycemia models is presented in

Table 5. Conceptual translational readiness categorization for ECG-based dysglycemia detection.

Maturity Level	General Characteristics	Validation Status	Dataset Requirements	Translational Meaning
Level 1	Proof-of-concept/exploratory study	Internal only or absent	Small, highly selective cohorts	Technical feasibility only
Level 2	Initial model development	Cross-validation/train–test split	Single-center datasets	Early methodological evidence
Level 3	Clinical ML validation	Independent test set, structured retrospective evaluation	Larger clinical datasets	Moderate translational potential
Level 4	Advanced clinical validation	External validation across cohorts	Multi-center/population-based datasets	Near-translational readiness
Level 5	Real-world deployment	Prospective and implementation evaluation	Representative screening populations	Clinically deployable screening system

.

In summary, achieving clinically viable ECG-based dysglycemia detection requires a shift from fragmented methodological approaches toward integrated, system-level solutions. The convergence of standardized multilead acquisition, physiologically informed feature extraction, robust machine learning, and rigorous validation provides a realistic pathway toward scalable and accessible screening systems. This perspective not only addresses the limitations identified in the current literature but also outlines a practical direction for future research and development. The integrated pathway from ECG acquisition to clinically deployable dysglycemia screening is illustrated in Figure 7.

This framework emphasizes that reliable screening cannot be achieved through isolated optimization of individual components, but requires coordinated development across the entire pipeline. Standardized multilead ECG acquisition ensures consistent and physiologically meaningful input data, while robust feature representation—combining data-driven and physiologically grounded approaches—enables extraction of clinically relevant patterns. These components must be coupled with machine learning models designed for generalizability rather than dataset-specific performance, and validated through rigorous external and prospective evaluation. Importantly, the conceptual pathway highlights that translational readiness is determined not by predictive accuracy alone, but by the alignment of data quality, methodological rigor, and system-level integration. This end-to-end perspective provides a practical foundation for advancing ECG-based dysglycemia detection from experimental studies toward scalable, real-world screening applications.

4.3. Clinical Implementation Considerations

Although ECG-based machine learning approaches for dysglycemia detection demonstrate promising methodological potential, their real-world clinical implementation remains challenging. Importantly, such systems are unlikely to replace established biochemical diagnostic methods such as fasting plasma glucose or HbA1c measurements. Instead, their most realistic near-term application may involve non-invasive pre-screening, population-level risk stratification, or identification of individuals requiring confirmatory laboratory testing.

From a public health perspective, the balance between sensitivity and specificity becomes particularly important in large-scale screening scenarios. High-sensitivity models may substantially increase false-positive referrals and unnecessary confirmatory testing, whereas overly specific models may fail to identify individuals with early-stage dysglycemia or prediabetes. Therefore, clinically meaningful operating thresholds should be determined according to the intended screening context and population characteristics rather than solely based on global performance metrics.

Wearable ECG technologies represent a promising direction for scalable and continuous monitoring applications. However, practical deployment remains constrained by signal quality variability, motion artifacts, acquisition heterogeneity, and the lack of standardized ECG acquisition and preprocessing protocols across devices and studies.

Overall, current evidence suggests that ECG-based artificial intelligence systems should presently be considered supportive screening or risk assessment tools rather than standalone diagnostic technologies. Further translational progress will require prospective validation, standardized workflows, and integration into realistic clinical screening pathways.

5. Limitations of This Review

This review has several methodological limitations that should be considered when interpreting the findings. Although the study was conducted following PRISMA-based principles, no formal review protocol was registered. The literature search was restricted to English-language publications indexed in major databases. Study selection and data extraction were primarily performed by a single reviewer with consistency verification, which, despite efforts to ensure accuracy, may introduce a degree of subjectivity compared to fully independent multi-reviewer procedures.

The synthesis is further constrained by substantial heterogeneity across the included studies, including differences in dataset characteristics, ECG acquisition protocols, feature representation, outcome definitions, and evaluation metrics. This variability precluded the use of quantitative meta-analysis and necessitated a narrative synthesis approach, which may be influenced by interpretative bias. Moreover, the proposed conceptual translational readiness categorization represents a conceptual integration of the reviewed evidence rather than a formally validated model, and its generalizability to broader clinical contexts remains to be established.

The substantial heterogeneity of included studies reflects the early-stage and interdisciplinary nature of the field. Consequently, the review was designed as a structured critical synthesis focused on methodological interpretation and translational analysis rather than formal comparative effectiveness assessment.

From a practical perspective, ECG-based dysglycemia detection is unlikely to replace established biochemical diagnostic methods such as fasting glucose or HbA1c testing. Instead, its most realistic clinical role may involve opportunistic pre-screening, population-level risk stratification, or continuous passive monitoring integrated into existing ECG workflows and wearable ecosystems. Potential use cases include preliminary screening during routine ECG examinations, wearable-based risk monitoring, remote health assessment in low-resource settings, and identification of individuals who may benefit from confirmatory biochemical testing.

6. Conclusions

This review synthesizes evidence from 17 studies investigating machine learning approaches for ECG-based detection of dysglycemia, highlighting both the promise and the current limitations of this emerging field. Across the reviewed literature, reported model performance is generally high, with accuracy and AUC values frequently exceeding 0.80 and in some cases approaching 0.95–0.99. However, these results are strongly dependent on dataset characteristics, as studies based on small or highly controlled cohorts tend to report substantially higher performance compared to those using larger and more heterogeneous populations. Only a limited number of studies employ large-scale datasets (n > 2000) or external validation, indicating that robust clinical evidence remains scarce.

From a methodological perspective, the analysis reveals pronounced heterogeneity in ECG acquisition protocols, feature representation strategies, and machine learning models, with no consensus regarding optimal approaches. While both deep learning and feature-based methods demonstrate the ability to capture dysglycemia-related patterns, model performance is primarily driven by data quality, dataset scale, and validation strategy rather than algorithmic complexity. Importantly, the majority of existing studies remain at early to intermediate stages of maturity, with relatively few approaches approaching translational readiness.

Taken together, the findings indicate that ECG-based dysglycemia detection is technically feasible but not yet clinically mature. Advancing this field will require a shift from performance-driven model development toward standardized, system-level approaches, including large and representative datasets, reproducible ECG acquisition, hybrid feature frameworks, and rigorous external and prospective validation. The integration of these elements within an end-to-end pipeline provides a realistic pathway toward scalable, non-invasive screening systems capable of deployment in real-world clinical settings.