Measuring the Impact of AI on Report-Drafting Efficiency in Chest Computed Tomography Interpretation: Retrospective Analysis

Background: Artificial intelligence (AI), particularly deep learning, has shown promise in enhancing medical image interpretation and improving radiologists’ efficiency. In China, growing imaging demand and workforce shortages have placed increasing pressure on radiology services. However, evidence on the operational impact of AI on reporting efficiency remains limited. Objective: This study aimed to evaluate the effect of an AI system on radiologists’ reporting efficiency by examining changes in report-drafting time for lung nodule diagnosis in chest computed tomography (CT) images. Methods: We analyzed 185,044 chest CT reports from Beijing Anzhen and Tsinghua Changgung Hospitals (2018-2023) using a difference-in-differences design with nonequivalent comparison groups. Report-drafting time before, immediately after, and up to 2 years following AI implementation was compared, adjusting for radiologist gender, seniority, and years of working experience. Results: The pooled analysis showed a modest overall increase of 0.86 minutes (95% CI 0.14 to 1.57). However, this masked substantial heterogeneity between hospitals due to differing implementation timelines. In the first year after AI deployment, Tsinghua Changgung Hospital experienced a nonsignificant increase of 0.90 minutes (95% CI –0.28 to 2.08). In contrast, at Beijing Anzhen Hospital, the AI-assisted group exhibited an absolute reduction of 0.76 minutes by the first year and a further 1.83-minute reduction by the second year (an approximate 28% time saved vs baseline), while the control group remained stable over time. Using a difference-in-differences framework, this corresponded to a 2.66-minute relative improvement compared with the counterfactual trend (P<.001). Conclusions: AI-assisted lung nodule diagnosis may initially increase report-drafting time due to adaptation and workflow adjustment. Sustained, meaningful efficiency gains were heterogeneous and observed at only 1 of the 2 study sites, indicating that long-term impacts are strongly contingent on site-specific implementation dynamics, learning curves, and local context.