Prediction of longitudinal clinical outcomes of MCI population using Alzheimer’s AT[N] biomarker profiling: Can machine learning methods improve our predictions?

• Published in: Alzheimer's & dementia Vol. 16
• Main Authors: Ezzati, Ali, Harvey, Danielle J., Davatzikos, Christos, Truelove‐Hill, Monica, Sreepada, Lasya P, Pomponio, Raymond, Zammit, Andrea R, Jack, Clifford R, Aisen, Paul S, Lipton, Richard B
• Format: Journal Article
• Language: English
• Published: 01.12.2020
• ISSN: 1552-5260; 1552-5279
• DOI: 10.1002/alz.041125

Background The NIA‐AA Research Framework on Alzheimer’s disease (AD) represents an important advance in the biological characterization of the AD spectrum. While this framework predicts disease progression, definitions of amyloid, tau, and neurodegeneration positivity based on cut‐scores may limit the performance of predictive models. Method 252 cognitively normal (CN), 424 amnestic mild cognitive impairment (aMCI) and 128 AD participants were included in this study. aMCI participants at baseline were classified using 2 different methods: 1) Using conventional AT[N] criteria based on proposed cut‐offs in CSF or brain imaging measures: participants were classified on the AD continuum (A+T‐N‐, A+T+N‐, A+T‐N+, A+T+N+) and non‐AD continuum (all other subgroups); and 2) using a machine learning (ML) based method (ensemble linear discriminant model built using CN and AD participants) to classify MCI participants as CN‐like or AD‐like based on continuous CSF biomarkers, regional amyloid PET, FDG‐PET, and volumetric MRI (variables contributing to AT[N] classification; Table 1). Predictive performance of AT[N]‐based classification and ML‐based classification were assessed using clinical outcomes at 0.5, 1, 2, 3, 4, and 5 years of follow‐up. Result A total of 414 aMCI participants had at least one wave of follow‐up (N of 397, 400, 346, 297, 238, and 93 at 0.5, 1, 2, 3, 4, and 5 years, respectively). The progression rate from aMCI to AD was 5.8% at 6m, 9.5% at 1y, 17.9% at 2y, 19.9% at 3y, 20.2% at 4y, and 15% at 5y of follow‐up. Presence of AD pathology at baseline (AT[N]‐based AD continuum) predicted progression to AD with high sensitivity but modest specificity during follow up (Table 2). ML‐based classification was less sensitive but substantially more specific. ML‐based classification models were significantly more accurate than AT[N]‐based classification at all follow‐up periods (McNemar test, p‐value<0.001 for all). Conclusion A single cut‐point approach used in AT[N] classification lacks accuracy when research questions require high diagnostic or prognostic certainty. A multivariate approach using ML methods can improve accuracy of prediction of disease progression. PPVs were higher at all follow‐up time points for ML‐based classification, indicating that progression to AD occurs at a higher rate in persons identified by ML.