Johns Hopkins community hospital reduces length-of-stay by 12+ hours with ML-based discharge prediction integrated into multidisciplinary rounds

Manual clinician-driven discharge predictions during morning huddles and multidisciplinary rounds were time-consuming, highly variable, and inaccurate. Avoidable delays near discharge exposed patients to hospital-acquired risks and blocked downstream access, contributing to ED boarding, PACU boarding, and OR holds. The hospital needed a scalable, automated approach to predicting individual patient discharges in real time.

Unit-specific random forest models were trained on EHR data from four inpatient units (two medical, one surgical, one telemetry) and deployed to generate discharge predictions at three daily time points per patient. Predictions were surfaced directly in the Epic EHR patient track board and via automated email. Case managers used the predictions as the opening step in multidisciplinary rounds, prompting structured team discussion to agree on a projected discharge date and prioritize discharge-blocking tasks.

Across 12,470 patients and 120,780 prospective predictions, AUC ranged from 0.70 to 0.80 for same-day and next-day discharge predictions. Implementing the tool reduced median hospital length-of-stay by more than 12 hours on the primary medicine unit (p<0.001) and on the telemetry unit (p=0.002). No statistically significant change was observed on the surgery unit or second medicine unit, highlighting variation in execution across teams.

• ML discharge predictions are most impactful when embedded directly into existing clinical workflows (multidisciplinary rounds) rather than delivered as standalone tools.
• Unit-level model customization is important — a single hospital-wide model would have masked meaningful variation in predictors and performance across care settings.
• Change management and just-in-time education (132 hours across units) were as critical as model accuracy; units with stronger rounding discipline saw the largest LOS reductions.