PHASE-ED: Recorded Emergency Department Stay and Temporal Prediction Across Pandemic Phases at Jordanian Public Hospitals

Abstract Shorter recorded emergency department stays may reflect changes in care, patient composition or documentation. PHASE-ED characterised recorded stays across pandemic phases, examined recording-period sensitivity and assessed prediction using sparse variables in Jordanian Ministry of Health records. Following a facility and coverage audit of \(\:\text{2017\--2025}\) data, the primary retrospective series comprised \(\:\text{382,018}\) eligible encounters from \(\:\text{85,356}\) patients at one facility during \(\:\text{2018\--2025}\); seven other facility codes contributed separately bounded analyses. Median recorded length of stay was \(\:\text{63.0}\) minutes before COVID, \(\:\text{60.0}\) during COVID and \(\:\text{51.0}\) during recovery. Corresponding \(\:\text{90}\)th percentiles were \(\:\text{317.0}\), \(\:\text{294.0}\) and \(\:\text{156.6}\) minutes, and proportions exceeding six hours were \(\:\text{8.33\%}\), \(\:\text{7.73\%}\) and \(\:\text{1.22\%}\). Recovery medians were already lower before an observed timestamp-precision transition on \(\:\text{23}\) July \(\:\text{2024}\). Historical and recent benchmarks, quantile regression and boosted trees were developed chronologically and evaluated on \(\:\text{191,497}\) later encounters. Test mean absolute error was \(\:\text{49.25}\) minutes for the historical benchmark, \(\:\text{47.66}\) for regression and \(\:\text{48.02}\) for boosted trees; the development simplicity rule retained the historical benchmark. Its nominal \(\:\text{80\%}\) interval covered \(\:\text{91.5\%}\) overall but only \(\:\text{68.7\%}\) within the recorded Resuscitation category. Thus, shorter recovery intervals coexisted with modest predictive gains and uneven uncertainty coverage.