Machine learning-based risk scores are associated with conversion to dementia in Veterans.

Area of Science:

• Neurology
• Gerontology
• Data Science

Background:

• Developed ancestry-specific risk scores for Alzheimer's disease and related dementias (ADRD) using machine learning (ML) and natural language processing on electronic health records of Black and White American (BA and WA) Veterans.
• Previous studies linked these scores to probable ADRD diagnoses at score generation, but future risk association was unclear.

Purpose of the Study:

• To evaluate the association between baseline ADRD risk scores and subsequent ADRD incidence and all-cause mortality.
• To assess the predictive utility of ML-derived risk scores in a diverse veteran population without prior ADRD diagnosis.

Main Methods:

• Conducted survival analyses on BA and WA Veterans without a prior ADRD diagnosis.
• Utilized cause-specific Cox proportional hazards models, treating death as a competing risk.
• Stratified analyses by race and conducted separately for BA and WA Veterans.

Main Results:

• Higher ADRD risk scores significantly correlated with increased risk of future ADRD diagnosis in both BA (HR=1.98) and WA (HR=2.13) Veterans.
• Elevated ADRD risk scores were also associated with increased mortality risk in both groups (BA HR=1.52, WA HR=1.55).

Conclusions:

• ML-derived ADRD risk scores are associated with increased risks of future ADRD development and mortality.
• These findings support the utility of ML-derived risk scores for early ADRD detection and prognosis in diverse populations.