Machine learning-derived biomarker cutoffs for Alzheimer's disease: Validation and application in preclinical and prodromal phases.

This study aims to derive biomarker cutoffs in Alzheimer's disease dementia (ADD) and validate their concordance in preclinical and prodromal stages. 341 sporadic and 103 familial participants were selected from the China Cognition and Aging Study (COAST) and the Chinese Familial Alzheimer's Disease Network (CFAN) cohorts, separately. Machine learning was used to generate prediction models and biomarker cutoffs for identifying ADD. Agreement test was used in cognitive normal and mild cognitive impairment (CN + MCI) participants. For COAST, the optimal regression model was CSF Aβ42/40, ptau and left medial temporal atrophy (MTA-L), with area under the curve (AUC) of 0.841. The optimal decision tree model was CSF Aβ42/ptau, Aβ42/ttau, and MTA-L (AUC = 0.820). For CFAN, the optimal regression model was left precuneus relative volume and MTA-L (AUC = 0.935). The optimal decision tree model was left hippocampal and precuneus relative volume (AUC = 0.806). They showed significant concordance in CN + MCI participants with cutoff-based diagnosis in ADD. Machine learning-enhanced thresholds could improve participant stratification in early Alzheimer's disease (AD) trials.