Effectiveness, Safety, and Workflow Burden of Large Language Model–Based Medical Report Generation: Systematic Review

Abstract

Abstract Background Systems based on large language models (LLMs), multimodal LLMs, and vision-language foundation models are increasingly being evaluated for medical report generation in imaging and related clinical workflows. Existing reviews have summarized technical architectures, radiology applications, readability, and benchmark performance, but clinical readiness remains uncertain because safety, human oversight, and workflow outcomes are sparsely and inconsistently reported. Objective The aim of this study is to assess the effectiveness (expert acceptance and blinded preference), safety (clinically significant, omission, and commission errors), and workflow burden (reporting time, corrections, edit distance, and editing burden) of LLM-based medical report generation. Methods We searched PubMed/MEDLINE, Embase, Web of Science Core Collection, Scopus, and the Cochrane Library for studies published from January 1, 2016, through May 15, 2026. Eligible studies evaluated LLMs, multimodal LLMs, or vision-language foundation models for image-to-report generation, impression generation from findings, report drafting, or structured reporting in imaging workflows. Two reviewers performed screening, extraction, risk-of-bias assessment, and Grading of Recommendations Assessment, Development, and Evaluation–informed narrative certainty assessment. Outcomes were clinically significant error rate, omission error rate, commission error rate, reporting time, edit burden, expert acceptance, and blinded expert preference. Meta-analysis was not performed because no comparable outcome had at least 2 studies with compatible task structure and analyzable data. Results A total of 101 studies were included. Chest x-ray was the largest modality group (36 studies), followed by computed tomography, magnetic resonance imaging (MRI), ultrasound, endoscopy, pathology, ophthalmic, electrocardiographic, dental, and mixed-modality contexts. No study was judged at low risk of bias; 15 were moderate, 72 high, and 14 serious. Safety and workflow evidence remained heterogeneous and largely nonpoolable. In a chest x-ray study, AI report acceptance was similar to that of radiologist reports (6047/8580, 70.5% vs 6288/8580, 73.3%), but false-negative findings were slightly higher (1584/8580, 18.5% vs 1527/8580, 17.8%). In a clinician-collaboration chest x-ray study, AI reports were equivalent or preferred in 233 of 300 (77.7%) and 170 of 303 (56.1%) cases across 2 datasets; yet, clinically significant errors persisted. In a brain MRI study, AI assistance reduced reading time from 61 to 53 seconds, whereas impression drafting increased editing time and edit distance. Conclusions This review shifts the synthesis from plausible report generation to clinically interpretable effectiveness, safety, and workflow effects. Expert acceptance and preference suggested assistive value in selected supervised settings, but these signals were limited by inconsistent reporting of clinically significant errors, omissions, commissions, and failed generations. Workflow effects were mixed, with some studies reporting shorter reading time or drafting support, and others reporting greater editing time or edit distance. The evidence remains too heterogeneous, biased, and sparse on case-level end points to support a pooled meta-analysis or autonomous clinical-readiness claims. Adoption should remain locally validated, clinician-supervised, and accompanied by standardized reporting of acceptance, preference, omissions, commissions, failed generations, reporting time, corrections, and editing burden.