Psychometric modeling of Boston process approach data for dementia prediction in the Framingham Heart Study

Abstract Objectives: Neuropsychological (NP) tests are multi-domain in execution. Reliance on a single score representing specific domains obscures the detection of subtle cognitive changes and increases risk of inaccurate assessment. Rooted in the Boston Process Approach (BPA), the Framingham Heart Study (FHS) captures multi-dimensional errors and process features within and across NP tests. We examined these BPA variables in community-dwelling older adults. Methods: We analyzed data from 2363 dementia-free participants aged 60 and above. Exploratory and confirmatory factor analyses used Kemeny covariance structures. Measurement invariance was estimated across age, sex, and education groups. We assessed the impact of demographics on latent factors, and the ability of these factors to predict future conversion to all-cause dementia. We trained machine learning (ML) models to compare NP and BPA data. Results: Participants were older adults (mean age 71.5 ± 8.7 years), primarily female (54.2%), and non-Hispanic White (96.5%). The bifactor model was the only model with adequate fit (CFI = 0.96, RMSEA = 0.03). General and specific factors captured ability for accurate and strategic responses, test-specific variance, and nuanced executive and semantic processes distributed across tests. Higher general ability and stronger verbatim story recall were associated with a reduced likelihood of developing all-cause dementia (general: OR = 0.15, 95% CI 0.12–0.86; recall: OR = 0.24, 95% CI 0.23–0.90) over a median of 5.2 years. With NP/BPA data, ML models identified >99% of 222 converters. Conclusions: This study highlights the strengths of NP/BPA data. Multidimensional cognitive features may enhance sensitivity to early changes predictive of incipient dementia.