Research Article Predicting 3-Year Mild Cognitive Impairment Decline in Community-Dwelling Older Adults: A Multimodel Machine Learning Study
by Ayuto Kodama1,2, Takako Ohnuma2, Kana Sasaki2, Kaoru Sugawara2, Nobuhiro Fujiyama4, Youko Umetsu5, Tsuyoshi Ono6, Hidetaka Ota2,3*
1Department of Occupational Therapy, Graduate School of Medicine, Akita University, Akita, Japan
2Advanced Research Center for Geriatric and Gerontology, Akita University, Akita, Japan
3Akita Prefectural Cardiovascular and Neurosurgical Center
4Clinical Research Promotion and Support Office, Future Cooperative Research Organization, Akita University, Akita, Japan
5Integrated Community Support Center, Public Health and Welfare Department, City Hall of Yokote, Yokote, Japan
6Omori Municipal Hospital, Yokote, Japan
*Corresponding author: Hidetaka Ota, PhD, Advanced Research Center for Geriatric and Gerontology, Akita University, 1-1-1 Hondo, Akita, Japan 010-8543.
Received Date: 18 March 2026
Accepted Date: 23 March 2026
Published Date: 26 March 2026
Citation: Kodama A, Ohnuma T, Sasaki K, Sugawara K, Fujiyama N, et al. (2026) Predicting 3-Year Mild Cognitive Impairment Decline in Community-Dwelling Older Adults: A Multimodel Machine Learning Study. Advs Prev Med Health Care 9: 1090. DOI: https://doi.org/10.29011/2688-996X.001090
Abstract
Background: Early identification of older adults at risk of mild cognitive impairment (MCI) decline is a public health priority given the rising global burden of dementia and the evidence that modifiable risk factors can be targeted before dementia onset. Community-based programs are well-positioned to screen at-risk individuals using routinely collected, low-cost assessments; however, reliable and well-calibrated prediction tools for short-term MCI decline in community settings are lacking. We developed and internally evaluated multivariable prediction models for 3-year MCI decline using baseline multidomain assessments from a community screening program. Methods: Community-dwelling adults aged ≥65 years underwent baseline demographic/health surveys, physical function testing, and cognitive assessments, and were followed for 3 years. The primary outcome was MCI change status (0= improve, 1= decline). We compared penalized logistic regression (glmnet), random forest, and gradient boosting (XGBoost), support vector machine, and k-nearest neighbors within a unified preprocessing pipeline. Performance was assessed on a stratified 70/30 train–test split using ROC-AUC, PR-AUC, sensitivity, specificity, balanced accuracy, and calibration metrics. Results: Of 266 participants, 79 (29.7%) experienced MCI decline at 3 years. In the held-out test set (n=81; decline n=24), penalized logistic regression achieved ROC-AUC 0.649 and PR-AUC 0.495. Calibration intercept was −0.624 and slope 0.360. Using a Youden/F1-optimized threshold (0.606), sensitivity was 0.458 and specificity 0.912; using a sensitivity-constrained threshold (≥0.70; 0.207), sensitivity was 0.708 and specificity 0.404. Conclusions: A parsimonious penalized logistic regression model using community-feasible predictors showed modest discrimination and imperfect calibration for 3-year MCI decline. External validation and model updating are warranted before clinical implementation; however, the approach highlights the feasibility of risk stratification using routinely collected measures in community settings.
Keywords: Mild cognitive impairment; Prediction model; Machine learning; Older adults; Community cohort; Calibration
© by the Authors & Gavin Publishers. This is an Open Access Journal Article Published Under Attribution-Share Alike CC BY-SA: Creative Commons Attribution-Share Alike 4.0 International License. Read More About Open Access Policy.