Dementia risk predictions from German claims data using methods of machine learning.

Area of Science:

• Health informatics
• Medical data analysis
• Dementia risk prediction

Background:

• German claims data are a rich source for epidemiological research.
• Predicting dementia risk is crucial for timely intervention and healthcare planning.

Purpose of the Study:

• To assess the suitability of German claims data for dementia risk prediction.
• To compare machine learning (ML) models against classical regression.
• To identify significant predictors of dementia risk.

Main Methods:

• Analysis of a 10-year follow-up dataset from 117,895 dementia-free individuals aged 65+.
• Inclusion of predictors: 23 age-related diseases, 212 prescriptions, 87 surgery codes, age, and sex.
• Application of logistic regression (LR), gradient boosting (GBM), and random forests (RFs).

Main Results:

• Moderate discriminatory power observed for LR (C-statistic=0.714) and GBM (C-statistic=0.707), with lower performance for RF (C-statistic=0.636).
• Gradient boosting models demonstrated the best calibration.
• Key predictors identified include antipsychotic medications, cerebrovascular disease, and a specific antibacterial prescription.

Conclusions:

• Dementia risk prediction models using German claims data achieve acceptable accuracy.
• These models offer potential for cost-effective decision support in early dementia screening.