Identifying Probable Dementia in Undiagnosed Black and White Americans Using Machine Learning in Veterans Health...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Geriatric Medicine

Background:

• Machine learning (ML) and natural language processing (NLP) in electronic health records (EHRs) show promise for reducing dementia underdiagnosis.
• However, ML models not reflecting minority populations may perpetuate health disparities, including dementia underdiagnosis.

Purpose of the Study:

• To develop and validate race-specific ML models for identifying undiagnosed dementia, with a focus on improving detection in Black Americans (BAs).
• To assess the performance of separate support vector machine (SVM) models trained on data from BAs and White Americans (WAs).

Main Methods:

• Developed SVM ML models using features from unstructured and structured EHR data to assign dementia risk scores.
• Trained separate models for BAs and WAs using electronic health records, including notes analyzed via latent Dirichlet allocation and stable topic extraction.
• Validated models on independent samples and used chart reviews by dementia experts to assess performance metrics like AUC, NPV, PPV, sensitivity, specificity, and accuracy.

Main Results:

• A strong positive relationship was found between SVM-generated risk scores and undiagnosed dementia.
• Black Americans were more likely than White Americans to have undiagnosed dementia (15.3% vs. 9.5% overall).
• The race-specific ML model for Black Americans demonstrated slightly better performance (AUC = 0.86) compared to the model for White Americans (AUC = 0.77) when validated against expert chart reviews.

Conclusions:

• Race-specific ML models derived from EHR data can effectively aid in identifying Black Americans at higher risk for undiagnosed dementia.
• These findings highlight the potential of tailored AI approaches to mitigate health disparities in dementia diagnosis.
• Future research should explore model generalizability across diverse populations (including more females) and clinical implementation strategies.