Predicting of β‐amyloid Positivity Using Integer‐based Risk Scores with Blood‐based Biomarkers

Abstract

AbstractBackgroundβ‐amyloid (Aβ), a pathologic hallmark of Alzheimer’s disease (AD) is the target of recently FDA‐approved drugs. In trials and in practice, identifying individuals likely to be β‐amyloid positive (Aβ+) with accessible and affordable tests will be increasingly important. Blood‐based biomarkers (BBMs) have shown early promise in determining amyloid status. We aimed to develop machine learning‐based integer risk scores using BBMs and other accessible measures to predict Aβ positivity by PET in older, non‐demented (cognitively unimpaired or mild cognitive impairment) adults.MethodsWe used data from 722 individuals (72.1 ± 7.0 years old) from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) to develop integer risk scores for predicting Aβ positivity. The sample was divided into training (75%) and test (25%) sets. Variables were categorized as demographics (D: age, sex, and education), genetic information (G: APOE4 status), cognitive measures (C: ADAS‐Cog‐13 and MMSE), and plasma biomarkers of p‐Tau181(T) and neurofilament light (N: NfL). Feature‐sets were formed as hierarchically constructed combinations of these categorized variables (see Figure 1). Using the AutoScore algorithm, variables were ordered according to their Random Forest Gini Index and were iteratively included in the construction of logistic regression‐derived risk score models. Variables were included based on increases in Area Under the ROC Curves (AUCs).ResultsFour models were developed based on four feature‐sets. All models utilized age and APOE4 status and, when considered, all models retained ADAS‐Cog‐13, p‐Tau, and NfL. The best performing model corresponded to the DGCTN feature‐set with performance on the dedicated test set of AUC = 0.810 (C.I.: 0.745‐0.876). Using the DGCTN model on the test set, 97% of individuals with scores in the top quartile of possible risk score points were correctly identified as Aβ+.ConclusionsWe developed risk score models utilizing accessible and affordable measures to predict Aβ positivity. Plasma biomarkers improve performance of risk scores indicating their value. Applications include screening in clinical settings for therapeutic care and enrichment of clinical trials. These data can be used to optimize the efficient use of PET in determining eligibility for clinical trials and amyloid targeted therapies in practice.