Wearables Flag Prediabetes—No Needles

Tablet with FDA on screen amid medical tools and ECG charts — Photo: ra2 studio / Shutterstock

The most important shift underway in prediabetes care is not a new drug or a fad diet, but a fundamental rethinking of how we find at-risk people in the first place—moving from needles and lab slips toward silent signals already embedded in the heartbeat.

At a Glance

An AI model called DiaCardia can identify prediabetes using electrocardiogram (ECG) data alone, including single-lead signals compatible with wearables.
Performance in retrospective studies is strong—area under the ROC curve around 0.85 across cohorts—but remains at the screening stage, not a replacement for diagnostic blood tests.
Current clinical guidelines still define prediabetes using glucose and HbA1c thresholds, and no society has yet endorsed ECG-based AI as a diagnostic standard.
The broader field of ECG-AI for diabetes shows promise but is constrained by generalizability, transparency, and limited prospective validation across diverse populations.

From Blood Draws to Heartbeats: What DiaCardia Actually Does

For decades, prediabetes has been defined at the laboratory bench: elevated fasting glucose, impaired glucose tolerance on the oral glucose tolerance test, or hemoglobin A1c in a narrow “intermediate” band below the diabetes threshold. These markers capture dysregulated glucose directly, but they require blood samples, trained staff, and a motivated patient willing to present for testing. DiaCardia approaches the same biological problem from an entirely different angle. Developed by a team at the Institute of Science Tokyo, the model ingests routine electrocardiogram recordings—either the full twelve leads used in clinical practice or the single lead approximating what a smartwatch sees—and uses machine learning to discern subtle changes in the electrical activity of the heart that correlate with prediabetes.

In their published work, the researchers report that DiaCardia distinguishes individuals with prediabetes from those with normal glucose status with an area under the receiver operating characteristic curve (AUROC) of roughly 0.85, a level of discrimination that compares favorably with many accepted risk scores. The same architecture was trained and tested across multiple health-check datasets, and the team emphasizes robustness across cohorts and independence from major clinical confounders such as age, sex, and body mass index. Critically, the model’s single‑lead variant—built on lead I, the vector often available from consumer wearables—retains much of this performance, opening the door to screening that could, in principle, run passively on devices people already wear every day.

How ECG-Based AI Finds a Metabolic Problem

The idea that a glucose disorder might be visible in an ECG is not as far-fetched as it sounds. Chronic hyperglycemia and insulin resistance influence myocardial structure and autonomic tone over time: conduction velocities, repolarization patterns, ventricular relaxation, and microvascular perfusion all shift as the cardiovascular system adapts—or fails to adapt—to metabolic stress. Traditional ECG interpretation is tuned to large, obvious changes such as bundle branch blocks or infarction patterns. DiaCardia and its predecessors ask a different question: if you treat each heartbeat as a rich, multidimensional signal and let a machine learning model examine thousands of such beats, can it detect a characteristic “fingerprint” of prediabetes that a human reader cannot?

Earlier work, such as the DiaBeats algorithm, showed that beat morphology-based machine learning could classify diabetes and prediabetes with high accuracy in retrospective datasets, reporting accuracies and F1 scores in the mid-90% range in an independent test set. Other groups have demonstrated deep learning models like IGRNet that diagnose prediabetes from short, twelve-lead ECG segments with strong reported accuracy and AUC in experimental cohorts. DiaCardia builds on this lineage but adds important refinements: interpretable modeling using methods such as SHAP value analysis to identify which ECG features drive predictions, systematic efforts to reduce confounding, and explicit attention to transportability across different screening sites. Mechanistically, the model is not claiming that ECG replaces glucose; it is leveraging the downstream cardiac manifestations of metabolic dysregulation as a non-invasive biomarker.

Where the Evidence Is Strong—and Where It Is Still Thin

It is important to be clear about what has been demonstrated and what has not. DiaCardia’s performance metrics are derived from retrospective analyses of health check cohorts in Tokyo, where participants had both ECG data and laboratory glucose measurements available. Pre-diabetes labels are defined using guideline-aligned criteria—HbA1c and fasting plasma glucose thresholds—so the model is, in effect, learning to approximate the output of the current diagnostic standard without drawing blood. Within that context, AUROC around 0.85 and high classification accuracy are meaningful. They suggest that, for this population, there is enough signal in the ECG to separate at‑risk individuals from normoglycemic peers.

Yet none of this converts DiaCardia into a diagnostic test in the regulatory sense. Contemporary guidelines from major organizations such as the American Diabetes Association and Canadian Diabetes Association still define prediabetes exclusively by laboratory criteria, with specific numeric cut‑offs for fasting glucose, HbA1c, and oral glucose tolerance results. External commentary, including endocrinology update forums, frames DiaCardia as a screening tool with “robust internal performance and external generalizability,” but explicitly notes that real-world impact will depend on prospective study design, confounding control, and outcome endpoints such as incident diabetes or cardiovascular events. The systematic review literature on ECG‑based AI for diabetes reaches a similar conclusion: strong potential, but constrained by heterogeneous methods, limited external validation, and a lack of standardized reporting that makes direct comparison across models and populations difficult.

Screening Versus Diagnosis: Why Guidelines Still Rely on Blood

Screening is about probability; diagnosis is about confirmation. A screening test’s job is to identify people who are more likely to have or develop a condition and should therefore undergo more definitive evaluation. In prediabetes, the definitive evaluation remains biochemical—glucose and glycated hemoglobin reflect the pathophysiology directly in a way regulators and clinicians trust and understand. For ECG-based AI to move beyond an adjunct role, two hurdles must be cleared. First, head‑to‑head studies in the same cohorts must show whether models like DiaCardia can match or improve upon blood tests for key endpoints: sensitivity, specificity, positive predictive value, and negative predictive value, not only for current prediabetes but for progression to overt diabetes and cardiovascular complications. Second, prospective, multicenter trials must demonstrate that using ECG screening changes what happens to patients in meaningful ways—catching disease earlier, reducing missed cases, and avoiding unnecessary referrals.

Those data do not yet exist at scale. Some AI-enabled ECG models for diabetes risk, such as the AIRE-DM system, have been trained to predict incident type 2 diabetes and show added value when integrated with clinical risk scores, suggesting that ECG features may improve long-range risk stratification. However, even these remain in the research domain, and no major guideline has incorporated ECG-AI as a primary pathway for diagnosis. Until robust prospective evidence arrives, the responsible stance is straightforward: use DiaCardia and similar tools as potential filters or alerts that prompt standard testing, not as substitutes for it.

Opportunities for Low-Friction, High-Reach Screening

Where DiaCardia’s promise is most compelling is not in replacing the phlebotomy chair, but in lowering the threshold for asking the question. Prediabetes is common, often asymptomatic, and strongly associated with both progression to diabetes and increased cardiovascular risk, including incident coronary events and heart failure. Many adults—particularly those who feel well—do not present for preventive lab testing, yet they undergo ECGs for unrelated reasons, from occupational health checks to pre-operative evaluations. If an AI model can scan those traces opportunistically and flag high‑risk individuals with reasonable accuracy, physicians gain a new way to identify patients who warrant further lab work, counseling, and follow‑up.

Wearable-compatible single‑lead models extend that logic further. Consumer devices capable of recording lead I ECGs are now widespread, and early technical work has shown the feasibility of synchronized multi-lead wearable systems that could, with proper validation, support richer signal analysis across glycemic states. DiaCardia’s architecture explicitly targets lead I, and institutional releases highlight “anytime, anywhere” screening as a plausible future. In resource-constrained settings, where laboratory capacity is limited or intermittent, an ECG-based pre-screen could help triage who most urgently needs biochemical testing. In high-income countries, it could add an unobtrusive layer to existing prevention programs, catching risk signals between clinic visits rather than waiting for fasting appointments.

Risks, Skepticism, and the Need for Careful Deployment

Skepticism around non-invasive AI biomarkers is not mere conservatism; it reflects hard lessons from prior waves of enthusiasm. Systematic reviews of AI approaches for non-invasive diabetes prediction underscore recurring issues: models trained on narrow, single-center datasets perform poorly when applied elsewhere; reporting is often inconsistent, with key details about calibration, class prevalence, and subgroup performance missing; and few studies provide transparent code or pre‑registered prospective validation. ECG-based prediabetes screening is no exception. DiaCardia’s strongest evidence comes from Tokyo health-check cohorts, and although the authors emphasize generalizability across sites, the public documentation still leaves questions about transportability to different ethnicities, healthcare systems, and device ecosystems.

Wearable deployment adds another layer of complexity. Consumer ECGs vary in hardware, signal quality, and noise characteristics; motion artifact, intermittent contact, and device firmware updates all influence the data a model would see. The pilot work on wearable 12‑lead systems explicitly stopped short of reporting diagnostic performance, noting that larger cohorts with biochemical confirmation are required before any screening claims can be made. Without rigorous device‑agnostic validation, there is a real risk of false reassurance—users who interpret a negative AI screen as a clean bill of metabolic health and delay conventional testing. The responsible path forward involves regulatory oversight, transparent communication that ECG-AI is a risk estimator rather than a diagnosis, and integration with clinical workflows that ensure abnormal results trigger, rather than replace, standard care.

What This Means for the Future of Prediabetes Care

Seen in the broader arc of cardiometabolic medicine, DiaCardia is best understood as the leading edge of a larger pattern: chronic diseases once defined solely by biochemical thresholds are increasingly being re-characterized as multi-system phenomena whose signatures can be read from organs not traditionally considered primary targets. The heart, in this case, becomes both victim and sensor. If ECG-based AI for prediabetes and diabetes continues to mature—through transparent methods, independent replication, and rigorous prospective trials—it could become an important component of layered screening strategies. The likely end state is not an either-or choice between ECG and blood tests, but a tiered approach in which inexpensive, ubiquitous signals identify who needs the more invasive workup.

For clinicians and health systems, the strategic question is how to experiment without overclaiming. Pilot programs that embed DiaCardia-like tools into existing health check infrastructure, with automatic reflex to laboratory testing for those flagged high risk, can generate real-world data while maintaining guideline-concordant care. Prospective studies in diverse populations—varying age, sex, BMI, ethnicity, and comorbidity—will determine whether AUROC figures from carefully curated cohorts translate into the messy reality of primary care. Meanwhile, patients should hear a clear message: this new way to screen for prediabetes does not remove the need for blood tests, but it may help ensure the right people get them sooner.