Leveraging Historical Patient Data to Identify Undiagnosed Diabetes and Prediabetes in Routine Care Free

Undiagnosed diabetes and prediabetes affect 8.7 million and 98 million American adults, respectively (1). Those with undiagnosed dysglycemia present with complications at diagnosis and disproportionately include adults who are younger, have obesity, belong to a race and ethnicity minority, or are uninsured (2–4). Even without progressing to a clinical diagnosis, individuals with undiagnosed diabetes or prediabetes face higher risks of atherosclerotic cardiovascular disease, heart failure, chronic kidney disease, and all-cause mortality, underscoring the urgency of early detection and intervention (5). Meta-analyses show that early pharmacological and lifestyle interventions for individuals with overweight or obesity, with or without prediabetes, can slow diabetes progression, reduce diabetes-related mortality, and improve cardiometabolic health (6,7).

Current guidelines recommend screening for dysglycemia every 3 years for asymptomatic adults aged ≥35 years with (U.S. Preventive Services Task Force) or without (American Diabetes Association [ADA]) overweight or obesity (7,8). While these guidelines are straightforward and practical for clinical use, they do not achieve high sensitivity and specificity (8). In our experience, most clinical diabetes diagnoses result from opportunistic screening based on patient risk factors rather than structured screening protocols.

As reported in this issue of Diabetes Care, Bowen et al. (9) developed and validated Dysglycemia Risk Score (D-RISK), a novel composite dysglycemia risk score designed to facilitate identification of undiagnosed dysglycemia (i.e., prediabetes and type 2 diabetes), measured as HbA_1c ≥5.7% (39 mmol/mol), using electronic health records (EHR) (9). D-RISK was developed with use of data from a retrospective cohort of 11,387 adults (18–64 years) who sought care between 2011 and 2014 at an outpatient primary care setting in Texas, and the approach was validated in a prospective cohort of 519 patients who were invited for screening between 2015 and 2017 at the same health care system. Both cohorts consisted of patients without diagnosed dysglycemia who had a valid random glucose test in the year prior to the index primary care visit. The score is additive and consists of categories of variables routinely collected in routine practice: age, race and ethnicity, BMI, random glucose, and history of hypertension. D-RISK attained an area under the receiver operating characteristic curve (AUC) of 0.75 (95% CI 0.74–0.76) in the derivation cohort and 0.71 (0.66–0.75) in the validation cohort for detecting dysglycemia. At a data-driven cut point of 9 (sensitivity 0.75, specificity 0.53, proportion eligible 56%), D-RISK outperformed the current screening guidelines of the ADA (AUC 0.52) and U.S. Preventive Services Task Force (AUC 0.58) and achieved higher discrimination than the ADA risk test (sensitivity 0.61, specificity 0.65, proportion eligible 44%, AUC 0.67). Across all tests, patients who are eligible for screening will be referred for diagnostic testing, consistent with standard clinical protocols (Fig. 1).

In addition to the interpretability and ease of implementation that D-RISK shares with the ADA risk test, a major advantage is the use of routinely available EHR data. Assuming the patient has previously used the health system and has a record of random glucose within the last 12 months, D-RISK does not require collection of additional data on variables such as physical activity or family history, unlike the ADA risk test. This enhances its scalability in using EHR data to facilitate risk assessment, since random glucose is often collected in metabolic panels in clinical practice. The discrete nature of D-RISK permits a range of screening thresholds with corresponding sensitivities and specificities, enabling health care providers to choose thresholds based on available resources and population risk levels. The adaptability is particularly beneficial in resource-limited settings. Second, the model was validated in a temporally distinct cohort within the same health care system. The consistent performance across both derivation and validation cohorts demonstrates the statistical generalizability of the approach.

Some limitations include that the derivation and validation sample consisted of users of a safety net health system serving predominantly vulnerable Hispanic individuals in Dallas, Texas. D-RISK needs to be validated in contemporary and non–safety net health systems. Additionally, since HbA_1c was the primary diagnostic tool for diabetes in this study, D-RISK may have missed some individuals with dysglycemia (e.g., isolated impaired fasting glucose) or misclassified those with conditions affecting its accuracy in the analytic sample (10).

A novel contribution of this work is the use of historical random glucose for prospective screening. Random glucose is routinely measured as part of metabolic panels but is not used in screening guidelines, given its intraday variability. Bowen et al. (11) previously showed that at the threshold of random glucose ≥100 mg/dL, one case of undiagnosed diabetes could be correctly identified for every 14 people screened. Rhee et al. (12) supported the predictive value of random glucose by demonstrating that two measurements of random glucose ≥130 mg/dL could detect type 2 diabetes within 1 year (sensitivity 59%, specificity 93%, AUC 0.90). Findings from this study suggest that random glucose, in addition to other variables, might serve as an early indicator of dysglycemia. Nevertheless, the independent contribution of random glucose warrants further investigation. Given the recent availability of over-the-counter continuous glucose monitors, future research should also explore how measures of glycemic variability can enhance dysglycemia screening.

The long-term impact of early diabetes screening, however, remains uncertain, as, in major trials, like Anglo-Danish-Dutch Study of Intensive Treatment in People with Screen-Detected Diabetes in Primary Care (ADDITION) and the Ely study, no significant improvements were found in population-level outcomes (13–15). In ADDITION-Cambridge, no reduction was found in all-cause, cardiovascular, cancer, or diabetes-related mortality after 10 years of follow-up (15). Similarly, while the Ely study showed that screening led to nearly 3-years-earlier diabetes diagnoses, no consistent improvements were found in clinical measures, medication use, or overall health outcomes, though a significant reduction in all-cause mortality was observed among individuals who were invited and screened (16,17). Besides, interventions for screen-detected diabetes, such as in ADDITION-Europe, have shown no significant impact on all-cause mortality or cardiovascular events (7,18). However, rather than negating the value of screening, these null findings highlight critical gaps that warrant further investigation. According to the classic screening criteria of Wilson and Jungner, effective screening must be linked to timely and accessible care beyond identification of individuals at risk (19). Many individuals with undiagnosed diabetes face structural barriers to health care and present with complications when diagnosed. They are also less likely to receive timely follow-up, thereby delaying interventions that could prevent long-term morbidity. Given that lack of insurance is a key characteristic of individuals with undiagnosed dysglycemia, broader access to health care needs to be addressed to truly close gaps in diagnosis. Furthermore, understanding the impact of different screening frequencies and optimizing care pathways after diagnosis remain pressing challenges.

EHR-based solutions can address these challenges in enabling systematic and scalable screening and prioritizing individuals with high risk for appropriate care. First, automated risk scoring, alerts, and patient registries can facilitate timely referrals and proper treatment. Second, longitudinal EHR data can augment current public health surveillance systems (20). Finally, machine learning and interpretable regression models using EHR data can predict diabetes 3–5 years before onset across diverse populations (21–25). Ultimately, prospective evaluations of whether these solutions can be effectively integrated into the clinical workflow for enabling disease prevention and screening are required before widespread adoption. D-RISK, therefore, offers a promising pragmatic solution to enable targeted diabetes screening with the potential for clinical translation.