Development and internal validation of a simple risk stratification score for unplanned emergency department revisits within 24 hours after discharge: A retrospective observational study

by Hidetaka Onda, Yuma Kuramatsu, Toyoaki Yamamoto, Atsunori Nishimatsu, Mami Kaneko, Daichi Tsutsui, Mizuki Kojima, Shinya Higuchi, Shinya Takeuchi, Kingo Nishiyama, Masato Miyauchi

Identifying patients at risk of very early emergency department revisits may help support discharge decision-making, but the extent to which routinely available index‑visit data can predict unplanned returns within 24 h remains unclear. Prior studies suggest that revisits are heterogeneous and influenced by factors beyond the initial encounter, underscoring the need to clarify—not overstate—the predictive value of simple, pragmatic tools.

We conducted a single‑center retrospective study of 2,774 emergency department discharge visits in 2024 in which blood tests were performed. The primary outcome was an unplanned revisit within 24 h to the same emergency department. Two physicians independently adjudicated revisits using administrative and chart data. Candidate predictors included arrival mode, initial vital signs, and routinely available laboratory variables. Continuous variables were screened across prespecified percentile‑based cutoffs, and a simple additive score was constructed from dichotomized items. Internal validation used 2,000 bootstrap resamples with full repetition of item selection and cutoff search. Calibration was assessed using the calibration slope, intercept, and Brier score.

There were 91 unplanned revisits (3.28%). Five dichotomized variables were retained: no ambulance use, C‑reactive protein ≥0.39 mg/dL, lymphocyte percentage ≤13.7%, red cell distribution width ≥13.5%, and mean corpuscular volume ≤87 fL. The odds ratio for revisit per 1‑point increase was 1.55 (95% confidence interval, 1.32–1.82). Apparent discrimination was modest (area under the curve, 0.651; 95% confidence interval, 0.592–0.711), with an optimism‑corrected area under the curve of 0.633. Apparent calibration in the derivation cohort was acceptable (slope 1.000; intercept 0.000; Brier score 0.0313). At a ≥ 4‑point threshold, specificity was 92.77% and sensitivity 20.88%.

A five‑item score based on index‑visit information showed a graded association with 24‑h revisit risk but modest predictive performance. This highlights the limitations of prediction using discharge‑time data alone and emphasizes the need for external validation and prospective evaluation before clinical implementation.