JMIR Publications

ABSTRACT

Background:

Machine learning is increasingly used in dementia risk prediction models, but their clinical utility and algorithmic fairness remain uncertain.

Objective:

This study aimed to develop and validate traditional statistical and machine learning models for dementia risk prediction in the general population and to evaluate their fairness across key subgroups.

Methods:

This population-based cohort study followed 483,824 UK Biobank participants without dementia for up to 13 years. Five models for all-cause dementia were developed using 56 evidence-based predictors: Cox proportional hazards model (Cox-PH), competing event risk model (CR), logistic regression (LR), XGBoost, and LightGBM. An age-only model was also developed for comparison. Model performance was assessed at various risk-stratified thresholds, and algorithmic fairness was evaluated through subgroup analysis.

Results:

During follow-up, 7,087 participants (1.46%) developed dementia. All models showed similar discrimination (C-index: 0.85-0.90). Although traditional statistical models (Cox-PH, CR, and LR) had lower C-indices than XGBoost and LightGBM in the internal validation cohort (C-index: 0.85-0.86 vs. 0.87-0.90), their performance in the external validation cohort was similar (C-index: 0.86-0.87 vs. 0.87). With 20% of the screening population identified as high-risk, detection rates ranged from 74% to 76%, with positive predictive values (PPV) of 5%, while the age-only model achieved 68% detection with PPV of 4% externally. The C index decreased to 0.59-0.70 in those aged over 70 years.

Conclusions:

Traditional statistical and machine learning models had comparable performance in predicting dementia risk within the general population. However, all models showed limited ability to predict high-risk individuals, especially in older populations, indicating limited clinical utility.