Accurate estimation of a hospitalized patient’s discharge date supports bed planning, throughput, and postacute care coordination. Machine learning applied to electronic health record (EHR) data can generate expected discharge dates,^1,2,3 and commercial tools are increasingly used in practice.^2,4 Most evaluations, however, benchmark a tool against the true discharge date in isolation rather than against the prevailing standard of care—the case manager’s estimate.⁵ We compared the accuracy of an artificial intelligence (AI)–generated discharge date with contemporaneous, independently generated case manager estimates and both estimates with the true discharge date across the hospital course.

This quality improvement study included inpatient encounters at Houston Methodist Hospital with discharge dates between August 1, 2023, and February 28, 2024. The hospital’s institutional review board determined the study exempt from review with a waiver of informed consent as data were deidentified. The study followed the SQUIRE reporting guideline. No baseline patient demographic data were available.

Artificial intelligence–estimated discharge dates were generated by a commercial clinician-assisted tool integrated with the EHR. Case manager–expected discharge dates were recorded during usual care in a workflow independent of the AI tool. Both were compared with the true discharge date from the EHR. The tool returned either an exact date or a categorical estimate; nonexact values were excluded, and 2.6% of encounters had no exact AI estimation (eMethods in Supplement 1).

We assessed estimations at 3 time points: admission (first available estimate per method) and 48 and 24 hours before discharge (first estimate in a symmetric window 36-60 and 12-36 hours before true discharge, applied identically to both methods). Accuracy was quantified as mean absolute error (MAE) and systematic bias as the mean signed difference (estimated minus true date), each with 95% CIs. Paired t tests compared methods (2-sided α = .05). Agreement was assessed using Bland-Altman analysis; Pearson correlation did not capture systematic bias between methods. Prespecified subgroup (length of stay [LOS], unit type) and sensitivity analyses were performed. Analyses were performed using R, version 4.3.3 (R Foundation for Statistical Computing) and Python, version 3.12.2 (Python Software Foundation).

A total of 22 349 inpatient encounters among 17 173 patients had both AI- and case manager–estimated discharge dates. The number per comparison varied by time point, since the near-discharge analyses required an in-window estimation (admission, n = 21 710; 48 hours, n = 9236; 24 hours, n = 12 172). Median (IQR) LOS was 5 (3-8) days. At admission, accuracy was similar between methods (MAE, 4.20 [95% CI, 4.08-4.31] vs 4.27 [4.15-4.38] days; P = .001), although case managers matched the exact date more often (23.6% vs 15.3%) and fell within 1 day more often (46.3% vs 40.8%). Nearer to discharge, case managers vs AI were more accurate (48 hours: MAE, 1.29 [95% CI, 1.26-1.32] vs 1.59 [95% CI 1.57-1.61] days; 24 hours: MAE, 0.98 [95% CI, 0.96-1.01] vs 1.93 [95% CI, 1.91-1.95] days) (both P < .001). At 24 hours, 79.5% of case manager estimations fell within 1 day vs 37.9% for AI estimations (Table; Figure).

Signed differences showed AI shifting from underestimating LOS at admission (mean difference, −2.74 [95% CI, −2.86 to −2.62] days) to overestimating it near discharge (mean difference, 1.24 [95% CI, 1.21-1.28] days at 24 hours), whereas case manager estimates remained close to the true date throughout (mean difference, −3.4 [95% CI, −3.53 to −3.28] and 0.04 [95% CI, 0.01-0.07] days, respectively). The 2 methods correlated strongly at admission (r = 0.92) but weakly nearer discharge (r = 0.12). Bland-Altman analysis showed a positive bias favoring later AI estimation, with wide limits of agreement at every time point (Table).

In subgroup analysis, relative accuracy varied by LOS, with AI less accurate than case managers for very short stays (1-2 days: MAE, 1.66 [95% CI, 1.62-1.70] vs 0.89 [95% CI, 0.84-0.94] days) but more accurate for intermediate stays (5-7 days: MAE, 1.49 [95% CI, 1.46-1.53] vs 2.24 [95% CI, 2.19-2.29] days), with both performing poorly beyond 14 days. Case managers were substantially more accurate on the obstetric unit, and both methods performed poorly in the intensive care unit. Accuracy varied widely across individual case managers. Restricting to 1 encounter per patient did not change the admission comparison.

This quality improvement study found that AI-estimated discharge was comparable to usual care at admission but outperformed by case managers as discharge approached (when accurate estimations are most actionable for bed management). The strong admission correlation did not persist nearer discharge and should not be interpreted as evidence that the tool captured the same information clinicians use throughout the stay. Any benefit seemed concentrated in intermediate-stay patients rather than uniform.

Several limitations apply. This single-center analysis assessed estimation accuracy, not downstream clinical decisions or patient outcomes; thus, the findings are hypothesis-generating. Case manager estimates were part of usual care and visible to care teams, so they may be partly self-fulfilling and, thus, advantaged nearer discharge; AI estimations were accessible in the EHR but not enabled automatically. Case management practices were not standardized, and performance varied across case managers. Whether the AI’s incremental accuracy in selected subgroups justifies the complexity of implementation warrants prospective, outcome-based evaluation.