Cross-Platform Concordance of Quantitative Amyloid PET Z-Scores in a Real-World Clinical Cohort of Patients with Cognitive Impairment and Suspected Alzheimer Disease.
Academic Article
Overview
abstract
BACKGROUND AND PURPOSE: Normative modeling tools are FDA-approved clinical software applications that enable quantitative amyloid PET analysis in routine clinical practice, however implementation differences may yield non-interchangeable Z-scores, with implications for diagnosis and treatment eligibility for patients with cognitive impairment and/or suspected Alzheimer disease. This study evaluated the agreement of Z-scores derived from two widely used clinical software packages in a real-world cohort of patients with cognitive impairment/Alzheimer disease. MATERIALS AND METHODS: Amyloid PET from 100 consecutive patients with cognitive impairment/Alzheimer disease obtained as part of standard-of-care evaluation at a single institution were retrospectively reviewed. Amyloid PET were post-processed using syngo.via MI Neurology (Siemens Healthineers) and MIMneuro (MIM Software/GE Healthcare). Regional Z-scores were obtained for the temporal, precuneus, posterior cingulate, parietal, frontal, and anterior cingulate cortices. Z-scores were compared per patient and stratified by Centiloid burden (low/intermediate/high: < 20/20-30/>30). Agreement was evaluated using Bland-Altman analysis and Deming regression. RESULTS: Agreement between Syngo.via and MIMneuro varied by region. When regional Z-scores were averaged into a per-patient composite measure, overall bias was near-zero, with tight limits of agreement (slope = 0.97 [95% CI 0.93-1.02], intercept = 0.11 [95% CI -0.05-0.26]), indicating minimal proportional and constant bias between platforms. Bland-Altman analysis showed small bias and narrow limits of agreement (LoA) in the low-centiloid group (mean bias -0.19, 95% LoA -0.93 to +0.55) and the greatest divergence in the intermediate group (mean bias -0.44, LoA -2.02 to +1.14). In the high-centiloid group, bias remained small (+0.16) with wider LoA (-1.92 to +2.24). Temporal cortex Deming regression demonstrated proportional and constant bias (slope = 0.70 [95% CI 0.66-0.74], intercept = 0.34 [95% CI 0.07-0.61]), indicating systematic underestimation of high Z-scores by MIM relative to Syngo.via. CONCLUSIONS: There was overall concordance between Syngo.Via and MIMneuro quantitative amyloid PET analysis software packages, however with significant region- and amyloid burden-dependent variability that increased in the intermediate-centiloid group. These differences may influence interpretation and determination of treatment eligibility in borderline cases, emphasizing that Z-scores from different commercial platforms should not be used interchangeably without cross-validation.
