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

Abstract

Background

We previously developed ancestry-specific risk scores for undiagnosed Alzheimer's disease and related dementias (ADRD) in Black and White American (BA and WA) Veterans by applying natural language processing and machine learning (ML) to Veterans Health Administration electronic health records. Using blinded manual chart reviews, we identified an association between ADRD risk scores and probable ADRD diagnosis at the time the scores were generated. However, it was unclear whether these scores were associated with future ADRD diagnoses and mortality.

Objective

To evaluate whether ADRD risk scores are associated with subsequent ADRD incidence and all-cause mortality among BA and WA Veterans without a prior ADRD diagnosis.

Methods

We conducted survival analyses to assess the association between baseline ADRD risk scores and time to either ADRD diagnosis or death. Cause-specific Cox proportional hazards models, treating death as a competing risk, were used to estimate hazard ratios (HRs) and 95% confidence intervals (CIs). Analyses were stratified by race and conducted separately for BA and WA Veterans.

Results

Higher ADRD risk scores were significantly associated with increased risk of developing an ADRD diagnosis (HR = 1.98, 95% CI: 1.72–2.27 for BAs; HR = 2.13, 95% CI: 1.79–2.54 for WAs) and mortality (HR = 1.52, 95% CI: 1.40–1.65 for BAs; HR = 1.55, 95% CI: 1.42–1.69 for WAs).

Conclusions

In addition to identifying undiagnosed cases, ML-derived ADRD risk scores are associated with increased risks of developing future ADRD and mortality, which supports their potential utility for both early detection and prognosis.