CogStrat: web-based tools for predicting incident cognitive impairment in cognitively normal older adults

Early identification of individuals at increased risk of cognitive impairment is crucial for timely intervention. We developed and evaluated Cox proportional hazards models and translated them into web-based applications to predict incident cognitive impairment in cognitively normal older adults using clinical data, with optional biomarker inputs. Data were drawn from cognitively unimpaired participants in the Alzheimer's Disease Neuroimaging Initiative (ADNI). Models included demographic and neuropsychological variables, with optional genetic, neuroimaging, and fluid (plasma and CSF) biomarkers. Performance was assessed using the C-index, calibration slope, and decision curve analysis, with internal validation via bootstrap resampling. Key models were externally validated in the Harvard Aging Brain Study. The ADNI sample comprised 729 participants (median age 73 years; 45% male). A clinical model including age, education, Montreal Cognitive Assessment, and delayed word recall showed acceptable discrimination (C-index 0.721; CI 0.672-0.772) and strong calibration (slope 0.940). Adding biomarkers improved performance, with the highest discrimination observed for a model including FDG and florbetapir PET uptake (C-index 0.805; CI 0.759-0.886). Discrimination remained stable after internal and external validation, while calibration was weaker for most plasma biomarkers. Five web applications (CogStrat) were developed using clinical data alone or combined with hippocampal volume, CSF tau/Aβ42, plasma neurofilament light chain, or FDG and florbetapir PET. These tools provide individualized risk group classifications, 5-year risk estimates, and 10-year cumulative incidence curves. CogStrat translates prognostic models into freely accessible, interactive tools that support risk stratification of cognitive impairment in cognitively normal older adults.