Using neuroimaging biomarkers to define plasma thresholds in a community dwelling cohort

Abstract

AbstractBackgroundAD Imaging biomarkers (IBM) include measures of amyloid (A) on PET, and neurodegeneration (N) on MRI, with thresholds for classification of imaging A/N positivity commonly used at research centers. Novel AD plasma biomarkers (PBM) can potentially enhance neuropathological characterization; however, knowledge regarding which thresholds to use to characterize PBM abnormalities is limited.MethodWe evaluated a range of PBM thresholds based on (1) plasma data alone (gaussian mixture models [GMM]), and (2) when classifying A/N IBM using both existing and data‐driven cutoffs (ROC analyses). Baseline plasma and imaging data were obtained in a community‐dwelling cohort enrolled in the Wake Forest ADRC, including cognitively normal participants (N = 300) and individuals with consensus diagnosis of mild cognitive impairment (N = 192) or dementia (N = 64; Table 1). We examined PBMs (Quanterix SIMOA HD‐X from NCRAD: Aß42/40, GFAP, NfL, p‐tau181) and IBM measures of A (global PiB PET; A+>1.21 SUVR) and N (FreeSurfer temporal cortical thickness; MAYOTHCK), N+ = 2.68mm; FreeSurfer hippocampal volume, N+ = 0.454% head size; HPC‐VOL).ResultOptimal/low/high GMM‐derived thresholds based on PBM distributions (Figure 1a) are provided in Figure 1b. ROC analysis using existing IBM cutoffs are in Figure 1c, with details (sensitivity/specificity/AUC) in Figure 1d. GMM‐ and ROC‐derived p‐tau181 thresholds were highly aligned (range 3.698‐4.071 pg/ml). For NfL and GFAP, imaging ROC‐derived thresholds were lower than GMM‐derived thresholds. AUCs were highest when PBM classification was used to predict A‐PET using existing PET‐positivity cutoffs, with PBM performing more poorly for MRI‐based cutoffs. A/N IBM histograms and current/optimal/low/high GMM‐derived thresholds for each IBM are in Figure 2a. Figure 2b represents how different data‐driven IBM cutoffs impact PBM thresholds in ROC analyses, with AUC statistics for these ROC analyses in Figures 2c‐e. Notably, GMM‐optimal IBM cutoffs performed similarly to current cutoffs for defining PBM thresholds, especially when using A‐PET (AUCs for current and GMM‐optimal >0.747).ConclusionWhile use of existing and data‐driven IBM cutoffs to define PBM thresholds in a community‐dwelling cohort is feasible, some plasma assays (e.g., p‐tau181) are more consistent than others, and caution is warranted against selecting a single threshold. To facilitate clinical use, future work will focus on establishing confidence intervals for PBM abnormalities.