AI-assisted radiographic fracture detection and length of stay in the adult ambulatory orthopedic emergency department: a before-after cohort study with a disease-specific internal control.

BACKGROUND: Although artificial intelligence-assisted radiographic fracture detection tools (AI-RFDT) have demonstrated high diagnostic accuracy in adult limb radiography, real-world evidence regarding their operational impact in the adult emergency department (ED) remains limited. METHODS: We evaluated whether deploying an AI-RFDT reduced ED length of stay (LOS) and revisit rates in fracture-excluded (F-E) patients undergoing limb radiography, using fracture-confirmed (F-C) patients as a disease-specific internal control. This retrospective controlled before-after cohort study, with complementary interrupted time-series analysis, was conducted in a tertiary Ambulatory Orthopedic ED. We included 8253F-E and 6864F-C visits between January 2021 and February 2026. Fracture status was assigned from the treating physician's ICD-coded discharge diagnosis, without study-level re-adjudication against imaging. The primary outcome was ED LOS before versus after deployment on 1 October 2023; secondary outcomes were 24-hour, 7-day and 30-day revisits. Subgroup analyses were exploratory. RESULTS: Mean F-E LOS decreased from 135.6 to 127.5 min (Δ -8.13 min; 95% CI -11.52 to -4.70; p < 0.001) and median LOS from 114.0 to 109.8 min (p < 0.001); the F-C control was unchanged (Δ + 0.37 min; 95% CI -3.95 to +4.74; p = 0.87); the cohort-by-period interaction was -7.99 min (95% CI -13.36 to -2.62; p = 0.004). The reduction was concentrated at the upper tail: no change at the 25th percentile and -24.0 min at the 90th (90th-25th difference -24.0 min; 95% CI -35.4 to -15.1); stays over four hours decreased from 11.3% to 8.2%. Interrupted time-series analysis showed a 19.2-min level decrease at deployment against a flat pre-existing trend, with no significant control change. Early revisits did not increase. CONCLUSIONS: AI-RFDT deployment was associated with decreased ED LOS in F-E patients, concentrated in the distribution's slow tail. The unchanged F-C control is consistent with an F-E-specific mechanism, with no increase in early revisits. These associations do not establish causation.