Reproducibility Before Treatment Decision Rules: A Targeted Scoping Review of AI-Assisted Coronary Plaque Characterization

Background/Objectives: Coronary computed tomography angiography (CCTA) can quantify plaque, whose burden predicts myocardial infarction better than stenosis severity. We reviewed evidence for artificial intelligence (AI)-assisted plaque characterization. Methods: This targeted scoping review (PRISMA-ScR) used a non-exhaustive search, mainly of MEDLINE/PubMed (July 2026, supplemented in September). AI title selection (135 of 1415 titles) and author assessment of 141 records yielded 121 included records under the revised criteria. An AI re-screen of 1280 unselected records (blind second pass on 175) found 110 further eligible records after charting; two AI passes screened 3634 confirmatory records (κ 0.93). The corresponding author accepted AI recommendations after reviewing written assessments, not blind (for the 335 confirmatory records, in one instruction). Of the 543 eligible records, 236 were charted (121 + 110 + 4 + 1; full text for 121 of the 236, abstract alone for 115) and 307 mapped from abstracts only (not fully charted). Results: In symptomatic or referred patients, AI-derived plaque burden was associated with major adverse cardiovascular events (a property of the measurement, not evidence that AI improves outcomes). In one comparison, AI modestly improved discrimination over manual quantification (AUC 0.82 vs. 0.78) without significant net reclassification improvement. No randomized trial of AI-based plaque analysis reporting clinical endpoints was identified in the sources searched and assessed; TRANSFORM (NCT06112418) is recruiting. Three small same-patient comparisons found systematic between-platform differences; no adequately powered comparison of major commercial platforms was identified. Abstract-level mapping of the 307 additional records identified no randomized outcome, observed decision-impact or photon-counting prognostic study. Conclusions: AI-derived plaque measures carry prognostic information in symptomatic patients, but randomized evidence that AI-guided treatment reduces myocardial infarction or cardiovascular death in asymptomatic adults, or does so better than coronary artery calcium (CAC)-guided prevention, has not been established in the sources searched and assessed. Under the modeled effects, demonstrating benefit in the lowest-burden stage would need an impractically large trial; as a design proposal, we outline a stage-enriched randomized trial. AI-assisted plaque characterization is best understood as a throughput and within-system repeatability technology; cross-platform portability is not demonstrated, and it is not a validated treatment decision rule.