Type 2 diabetes and prediabetes remain substantially underrecognized, highlighting the need for practical screening approaches based on routinely available data. This study aimed to develop and temporally validate a low-cost complete blood count (CBC)-based machine learning model for identifying individuals at elevated risk of diabetes in a routine health examination setting. We conducted a retrospective single-center study using electronic health record data from Shenzhen University General Hospital. A development cohort of 70,725 individuals examined between 2018 and 2023 was used for model development with stratified 5-fold cross-validation, and an independent temporal validation cohort of 26,650 individuals examined in 2024 was used for final evaluation. The outcome was defined as a binary diabetes-risk label based on HbA1c, with HbA1c < 5.7% classified as non-risk and HbA1c ≥ 5.7% classified as diabetes-risk, including both prediabetes and HbA1c-defined diabetes. Candidate models included logistic regression, LightGBM, multilayer perceptron, Tabular ResNet, and an attention-based tabular neural network. In the development cohort, LightGBM achieved the best overall performance, with a mean AUROC of 0.821 ± 0.006 and a mean AUPRC of 0.628 ± 0.010. In the independent 2024 temporal validation cohort, LightGBM again performed best, with an AUROC of 0.791 and an AUPRC of 0.699. Calibration analysis showed preserved risk ranking but some overestimation of absolute risk, and subgroup analysis revealed performance heterogeneity across sex, age, and BMI strata. These findings suggest that CBC-based machine learning, particularly LightGBM, may provide a practical approach for opportunistic identification of individuals at elevated risk of diabetes and support further confirmatory testing in routine health examination settings.
Guo et al. (Fri,) studied this question.