Understanding the impact of PET amyloid cutpoints on prognostic modelling for cognitively normal individuals

Abstract

AbstractBackgroundAmyloid beta (Aβ), measured using PET imaging, is a key biomarker for Alzheimer’s disease (AD) with the consequences of abnormally high Aβ levels (Aβ+) well‐established from prospective research cohorts. A critical question is whether the prognostic capabilities of Aβ can be improved further, for example by refinement of optimal criteria for abnormality.To date, existing studies have explored such issues using association analyses, which may not reflect performance in prognostic settings due to potential overfitting. Here, the impact of different Aβ cut‐points is determined in a cross‐validation framework, providing performance estimates on data from individuals that were not used for model construction, which better reflects realworld prognostic application. Using data for cognitively normal individuals (CN) from ADNI and AIBL, we estimate time to i) MCI or AD diagnosis and ii) cognitive deficit, defined as MMSE≤26.MethodWe analyse measurements from 344 and 748 CN from ADNI and AIBL respectively who have available PET Aβ scans. PET Aβ SUVRs were transformed to the centiloid scale (CL). For each task, the Aβ cut‐point is varied from ‐10 to 65CL and Cox models are constructed within 10 repeats of 10‐fold cross‐validation. From the resulting 100 models, performance is quantified as the median concordance index (i.e. Harrell’s C).ResultDetails of the two cohorts are shown in Table 1. Across both AIBL and ADNI, a PET only model shows robust performance for cut‐points within a wide range (5 and 50CL) for predicting either time to diagnosis cognitive deficit (Figure 1), with performance dropping rapidly outside this range. When additional covariates are included 2, we see maximal performance for lower cutpoints (5‐20CL) for diagnosis in ADNI and cognitive deficit in ADNI, while remaining tasks show improved performance with higher cut‐point ranges (20‐50CL). Trends in cut‐point are consistent regardless of covariates. Leaving Aβ as a continuous variable yields near‐optimal performance across all tasks.ConclusionOur results suggest that within a range (5 and 50CL), prognostic performance is robust to the choice of cut‐point for Aβ, suggesting further refinement of a single cut‐point within this range may not yield substantial improvements for prognostic tasks for CN individuals.