Diabetes mellitus is the eighth leading cause of death in the United States and contributes to substantial health care costs [1]. In 2021, an estimated 38.4 million Americans of all ages had diabetes, representing 11.6% of the US population [2,3]. Of those with diabetes, 22.8% were unaware of or did not report having diabetes [2,3]. When diabetes is undiagnosed, and consequently hyperglycemia remains unmanaged, severe and irreversible microvascular and macrovascular complications can develop, including diabetic neuropathy, nephropathy, retinopathy, and cardiovascular disease [4-7].

Screening asymptomatic individuals for undiagnosed diabetes enables earlier diagnosis and treatment, ultimately reducing the risk of complications and premature death [8-10]. The latest American Diabetes Association (ADA) and US Preventive Services Task Force guidelines recommend beginning diabetes screenings at the age of 35 years [11,12]. However, diabetes screening guidelines that rely on blood testing are not widely followed. Only 50%‐60% of US adults who met the criteria for screening reported receiving glucose testing within the past 3 years [13,14]. The testing rate was alarmingly low among high-risk groups, including those with low education, low household income, and limited health care access [13,14].

Risk assessment tools for diabetes detection using easily accessible and self-reported data have been proposed, but they have shown low overall accuracy and validity in the general population [15-21]. In recent years, various machine learning algorithms have been used to predict diabetes and have yielded better performance than traditional statistics-based models [22-28]. Few studies have developed machine learning models to detect undiagnosed diabetes in the US population [29,30]. Although 2 studies reported a good overall accuracy of 80%, the quality of a positive prediction by models (ie, precision) was notably low, which could lead to a high number of false positives and unnecessary follow-up testing [29,30].

More recently, there has been growing interest within the health care community in automated machine learning (AutoML), which automates machine learning models’ selection, composition, and parameterization to optimize performance [31-33]. AutoML uses voting and stacking ensemble techniques to combine multiple learning models, often improving classification accuracy more effectively than a single machine learning algorithm. Its automation also reduces human error and bias by impartially exploring a wide range of machine learning models [33]. However, despite its potential, no prior studies have investigated the feasibility and performance of AutoML in screening for undiagnosed diabetes.

This study aimed to investigate the potential use of AutoML and self-reported data in detecting undiagnosed diabetes among US adults in a nationally representative survey. The trained model could aid in detecting undiagnosed diabetes in the general US population, particularly in underserved populations with limited access to blood glucose tests. This study could also promote the adoption of AutoML in diabetes research.

In total, the study cohort included 11,815 participants, with 9559 participants without diabetes and 2256 with undiagnosed diabetes. The characteristics of the study cohort are summarized in Table 2. The average ages were 59.76 (SD 15.0) years for those with undiagnosed diabetes and 46.78 (SD 17.2) years for those without. The diabetes group had a higher proportion of males, lower levels of education, and a greater likelihood of having a relative with diabetes. Additionally, patients with undiagnosed diabetes had higher BMI, larger waist circumference, and a higher prevalence of chronic conditions. The flow diagram of participant selection is presented in Figure 1.

The performance of the AutoML model and traditional machine learning models is summarized in Figure 2. The AutoML model demonstrated superior performance compared to the 4 traditional machine learning models—logistic regression, support vector machines, random forest, and XGBoost. The trained AutoML model achieved an AUC of 0.909 (95% CI 0.897-0.921) and an accuracy of 86.5% in the test set. The model demonstrated a sensitivity of 70.26%, specificity of 90.46%, PPV of 64.10%, and NPV of 92.61% for identifying undiagnosed diabetes from nondiabetes (Table 3). The model summary and details were provided in Multimedia Appendix 1.

The top 5 features are age, waist circumference, daily total sugar intake, income, and BMI, together accounting for 50% of total model importance. Comorbidities, except hypertension, contributed minimally relative to demographic and behavioral factors. Excluding comorbidities (except hypertension) resulted in comparable model performance, with an AUC of 0.830 and an accuracy of 85.1%.

Additional analysis results are summarized in Table 4. The model using diabetes diagnosis criteria with a confirmatory test achieved a testing accuracy of 98.0% and 89.7% with an AUC of 0.823. However, precision (ie, PPV) and recall (ie, sensitivity) were suboptimal due to the small number of patients meeting the diabetes criteria with ≥2 tests. The model demonstrated a sensitivity of 44.10%, specificity of 92.22%, PPV of 24.23%, and NPV of 96.69% for identifying undiagnosed diabetes from nondiabetes. The performance of the multiclass prediction model was poor, with an overall accuracy of 58.9% for using diagnosis criteria with ≥1 test and 67.1% for using diagnosis criteria with ≥2 tests.

To our knowledge, this study is the first to utilize the AutoML model for detecting undiagnosed diabetes in US adults. The best-performing model achieved an AUC of 0.91 and an accuracy of 86.5% in the test set. National surveillance shows that nearly half of those with undiagnosed diabetes have hypertension, lipid abnormalities, or cardiovascular and chronic kidney diseases [45-48]. Delayed diagnosis of diabetes hinders the opportunities for early intervention to slow the progression of dysglycemia and its comorbidities. Our model was trained and tested using a substantial and diverse dataset comprising nationally representative survey data. The model’s high accuracy and applicability to the broader US population make it a promising tool for large-scale diabetes screening efforts.

The feature importance ranking of the best-performing model highlights waist circumference, BMI, and dietary variables as key predictors, underscoring their strong links to metabolic health, insulin resistance, and dietary habits that influence diabetes risk. Lifestyle factors such as drinking frequency and physical activity also emerged as significant contributors, whereas self-reported comorbidities played a smaller role once anthropometric and behavioral measures were included. These findings align with epidemiological studies and improve the model’s interpretability, providing actionable insights to prioritize targeted interventions for modifiable risk factors [49].

Machine learning has advanced clinical research but faces adoption barriers like data access, imbalances, and reliance on data science expertise for deployment [50,51]. AutoML reduces the need for machine learning expertise, enabling clinicians to use advanced technologies without programming skills and integrating them into research and clinical practice [52-55]. Despite being promising, few studies have explored the application of AutoML for diabetes diagnosis [56].

Previous studies have compared traditional machine learning models with conventional statistical models for identifying undiagnosed diabetes, demonstrating that machine learning models outperform statistical models [29,30]. These studies reported AUC values between 0.73 and 0.81, consistent with the performance of traditional models in this study. However, the reported low PPVs highlighted the limited ability of these models to accurately identify undiagnosed diabetes cases [30]. This study showed that AutoML models are superior and outperformed traditional machine learning models in detecting undiagnosed diabetes. Similarly, one study has reported that the AutoML model outperformed both individual and ensemble models in identifying patients with diabetes using electronic medical records data [57]. These findings suggest that AutoML provides a more accessible and efficient approach, eliminating the need for manual optimization while delivering superior performance.

Nonetheless, several issues with the AutoML model in diabetes screening should be noted. When applying more stringent diabetes diagnostic criteria, the accuracy reached 90%; however, precision and recall were low, likely due to the limited number of samples that met the ≥2 test criteria. H2O AutoML provides significant advantages, including its built-in class balancing functionality, which automates the handling of moderate class imbalance without requiring external implementations. Using random sampling to upsample minority classes or downsample majority classes, it effectively manages datasets with moderate imbalance, such as the 4:1 ratio under the 1+ test criterion for diabetes diagnosis in this study. However, with the stricter 2+ test criterion, the imbalance ratio rose to 24:1, likely exceeding H2O AutoML’s ability to mitigate the imbalance. This severe imbalance impacted the model’s ability to accurately distinguish undiagnosed diabetes from nondiabetes in unseen data, reducing precision and recall. For such highly imbalanced datasets, combining AutoML with other data resampling methods, such as Synthetic Minority Over-sampling Technique, could better improve model performance [58].

In addition, the model’s performance in multiclass classification of no diabetes, prediabetes, and undiagnosed diabetes was notably poor. Prediabetes is an intermediate stage between normal glycemia and diabetes and is highly prevalent [59]. Clinically, prediabetes and diabetes, as the continuum of dysglycemia, share many overlapping risk factors, such as insulin resistance and elevated glucose levels [60], making their differentiation challenging. The subtle metabolic differences between these conditions may not have been adequately captured by the included self-reported data. Future efforts should focus on incorporating additional features and refining the model architecture to enhance accuracy and improve its ability to identify prediabetes.

A major strength of this model is its use of NHANES data, which is nationally representative of the US population, enhancing the generalizability of the findings. The model incorporates comprehensive self-reported data, including nutritional information, for predicting undiagnosed diabetes. The application of AutoML in this study represents the first use of this approach in diabetes research, providing a foundation for further development and validation of similar models. However, the widespread adoption of our AutoML in health care requires further development and validation. False positives can lead to unnecessary tests, higher costs, and patient anxiety, while false negatives may delay treatment and worsen outcomes. Using AutoML models in real-world clinical settings requires meticulous threshold optimization and validation to achieve an appropriate balance between precision and recall. These models should be rigorously evaluated for fairness and equity, ensuring that performance does not vary significantly across demographic groups [61,62]. Practical barriers to implementing AutoML in clinical practice also include integration with electronic health record systems and the lack of trust in “black-box” models due to their opacity [31,61-63].

Several limitations should be noted. First, the definition of undiagnosed diabetes in the base model was based on a single elevated measurement, which may not fully capture the condition. However, the model still showed utility as a screening tool. Second, the model demonstrated poor performance in multiclass prediction, including prediabetes, indicating that additional feature refinement and model adjustments may be necessary to improve accuracy in 3-class predictions. Third, known diabetes diagnoses rely on self-reports, which may introduce potential recall bias and misclassification. This could lead to mislabeling cases and diminishing model performance. Although most published public health studies still rely on self-reports—hence making our study not an unusual one in using self-reports—future studies should aim to validate self-reported data against medical records where feasible to minimize errors and increase reliability of the observations. Finally, the generalizability of this model may be limited to US populations and may not extend to non-US populations.

This study demonstrates the potential of AutoML in detecting undiagnosed diabetes using self-reported and easily accessible data. Although challenges remain in accurately classifying multiple categories, including prediabetes, the model shows promise as a tool for large-scale diabetes screening. Further refinement and validation are required to improve its applicability across diverse populations.