Prediction of brain amyloidosis: a multi‐site machine learning analysis

Background Position emission tomography (PET) imaging and cerebrospinal fluid (CSF) analysis are the gold standards for evaluating brain amyloidosis in vivo, but their utility is limited by invasive procedures, high costs of PET, and limited accessibility to PET facilities. Prior research has used machine learning to assess amyloidosis based on non‐invasive and easily accessible information such as cognitive test scores and genetic data, but past accounts have typically used a single cohort to train and test their models, which may undermine their generalizability. Importantly, most studies have imputed amyloidosis of symptomatic participants without evaluating their models on cognitively unimpaired (CU) individuals. Accordingly, we aimed to develop algorithms that estimate risks for amyloidosis and validate them on CU participants from an external cohort. Method Data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) (n = 987) and the Charles F. and Joanne Knight Alzheimer Disease Research Center (Knight ADRC) (n = 688) were used. The examined predictors for amyloidosis were demographic factors, body mass index (BMI), apolipoprotein E (APOE) genotype, and performance on cognitive tests (Table 1). Random forest models were trained with nested cross‐validation (CV) on the ADNI cohort before testing on the Knight ADRC cohort for a reliable assessment of generalizability. The outer CV loop enabled unbiased measurement of prediction performance, and the inner loop hyperparameter optimization. Result The model achieved an area under the receiver operating characteristic curve (AUC) of 0.81 [0.79, 0.84] on the ADNI cohort (Table 2), and its performance generalized well to the Knight ADRC participants. When only CU participants are considered, performance was overall lower than that for the full cohort, which we expect given that they have less over pathology. The model was nonetheless able to achieve an AUC of 0.78 [0.75, 0.81] on the Knight ADRC CU participants. Conclusion Machine learning models can predict brain amyloidosis with a moderate degree of accuracy on asymptomatic individuals and generalize well to external cohorts. Our models can be used to screen for individuals likely to be Aβ PET‐positive, thus reducing preclinical AD trial recruitment cost by avoiding unnecessary PET scans.