Development and internal validation of exploratory prediction models for self-reported physician-diagnosed mild cognitive impairment and dementia among people with treated diabetes: a prediction modeling study using longitudinal cohort data from the KLoSA

Diabetes is a well-established risk factor for cognitive decline, yet tools for early identification of individuals at high risk of mild cognitive impairment (MCI) and dementia remain limited, particularly in Asian populations. Exploratory machine learning–based prediction models for incident MCI and dementia were developed and internally validated among individuals with diabetes using a nationally sampled panel. Data came from the Korean Longitudinal Study of Aging (2018–2024). Adults aged ≥ 45 years treated for diabetes and free of baseline cognitive impairment were included, leaving 1,246 for MCI and 1,255 for dementia. Incident MCI and dementia were ascertained from self-reported physician diagnoses and modeled as a six-year cumulative risk over a fixed prediction horizon from predictors measured at a single baseline wave, without repeated-measures modeling; the models predict self-reported diagnoses rather than standardized cognitive testing, and a sensitivity analysis used the Korean Mini-Mental State Examination at the final wave. Eight machine learning algorithms were evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, predictive values, and Brier score, with SHapley Additive exPlanations. During follow-up 33 participants (2.6%) developed MCI and 50 (4.0%) dementia. Fully nested cross-validation, the primary estimate, gave AUCs of 0.688 for MCI and 0.680 for dementia; repeated cross-validation at fixed hyperparameters gave higher values (0.709 and 0.719) that do not account for tuning optimism. The algorithm ranked first varied across repeats, so none can be identified as best. Single-split AUCs (0.854, 0.741) were most optimistic. Both showed high negative but low positive predictive values. Deaths were identified by exit-survey linkage: 23% died within six years. With death as a competing event, cumulative incidence was 2.8% for MCI and 4.2% for dementia; weighting for non-death dropout changed little. Key predictors were depressive symptoms, hearing difficulty, and education for MCI, and age, sex, and household income for dementia. These preliminary, internally validated models may support risk stratification but require external validation before clinical use. Given the small number of events, low positive predictive values, and self-reported outcomes, the findings are hypothesis-generating. The dementia model over-estimated absolute risk roughly ninefold; recalibration should precede external validation.