Comparison of Expert Reads Versus Local Clinical Reads of Amyloid PET Scans in the IDEAS Study.
TL;DR
Visual interpretation of amyloid PET can be performed in a real-world clinical setting with high reliability, and lower concordance between local and expert reads was associated with low local reader confidence.
Abstract
Amyloid-targeting therapies and growing reimbursement are expected to increase the clinical use of amyloid PET. The Imaging Dementia-Evidence for Amyloid Scanning (IDEAS) study captured more than 18,000 clinical scan interpretations from local radiologists and nuclear medicine physicians in the United States. This study aimed to evaluate the reliability of these interpretations by comparing local clinical reads of a representative sample of IDEAS scans to retrospective majority interpretations by expert readers. Methods: A representative subset of amyloid PET scans from the IDEAS study, 500 performed with each of the Food and Drug Administration-approved tracers (18F-florbetaben, 18F-florbetapir, 18F-flutemetamol) and previously interpreted by local readers, was assigned to vendor-selected panels of 3 expert readers (1 panel per tracer). The scan subset was stratified to match the overall study frequencies by participant age, level of cognitive impairment (mild cognitive impairment vs. dementia), positive or negative local interpretation, and PET facility type. Expert readers performed visual interpretations using tracer-specific criteria, blinded to clinical and demographic data. Cohen κ assessed agreement between local reads and the majority expert panel reads, and Fleiss κ measured overall agreement among expert readers. Results: Expert panels interpreted 1497 scans. Participants were a median age of 75 y (range, 65-96 y), 51.4% were female, and 60.5% had mild cognitive impairment. Local readers interpreted 60.9% of scans as positive and 39.1% as negative. Across all radiotracers, agreement between majority expert and local reads was good (κ = 0.76; 95% CI, 0.73-0.80), with 86.6% (790/912) agreement for positive scans and 91.5% (535/585) agreement for negative scans. Agreement did not vary significantly by PET tracer (κ = 0.73-0.78). Agreement between individual experts and local readers was also good (pairwise κ = 0.61-0.78), and agreement within the expert reader group was similarly high (κ = 0.79; 95% CI, 0.77-0.81). Lower concordance between local and expert reads was associated with low local reader confidence (P < 0.001). Conclusion: We found good agreement between local readers and the majority assessment of 3-member expert panels. Visual interpretation of amyloid PET can be performed in a real-world clinical setting with high reliability.