Comparing machine learning-derived MRI-based and blood-based neurodegeneration biomarkers in predicting syndromal conversion in early AD.

Abstract

We compared the machine learning-derived, MRI-based Alzheimer's disease (AD) resemblance atrophy index (AD-RAI) with plasma neurofilament light chain (NfL) level in predicting conversion of early AD among cognitively unimpaired (CU) and mild cognitive impairment (MCI) subjects. We recruited participants from the Alzheimer's Disease Neuroimaging Initiative (ADNI) who had the following data: clinical features (age, gender, education, Montreal Cognitive Assessment [MoCA]), structural MRI, plasma biomarkers (p-tau181 , NfL), cerebrospinal fluid biomarkers (CSF) (Aβ42, p-tau181 ), and apolipoprotein E (APOE) ε4 genotype. We defined AD using CSF Aβ42 (A+) and p-tau181 (T+). We defined conversion (C+) if a subject progressed to the next syndromal stage within 4 years. Of 589 participants, 96 (16.3%) were A+T+C+. AD-RAI performed better than plasma NfL when added on top of clinical features, plasma p-tau181 , and APOE ε4 genotype (area under the curve [AUC]=0.832 vs. AUC=0.650 among CU, AUC=0.853 vs. AUC=0.805 among MCI) in predicting A+T+C+. AD-RAI outperformed plasma NfL in predicting syndromal conversion of early AD. AD-RAI outperformed plasma NfL in predicting syndromal conversion among early AD. AD-RAI showed better metrics than volumetric hippocampal measures in predicting syndromal conversion. Combining clinical features, plasma p-tau181 and apolipoprotein E (APOE) with AD-RAI is the best model for predicting syndromal conversion.