Objective:
To develop a model to produce real-time, updated forecasts of patients’ intensive care unit length of stay using naturally generated provider orders. The model was designed to be integrated within a computerized decision support system to improve patient flow management.
Design:
Retrospective cohort study.
Setting:
Twenty-six bed pediatric intensive care unit within an urban, academic children’s hospital using a computerized order entry system.
Patients:
A total of 2,178 consecutive pediatric intensive care unit admissions during a 16-month time period.
Measurements and Main Results:
We obtained unit length of stay measurements, time-stamped provider orders, age, admission source, and readmission status. A joint discrete-time logistic regression model was developed to produce probabilistic length of stay forecasts from continuously updated provider orders. Accuracy was assessed by comparing forecasted expected discharge time with observed discharge time, rank probability scoring, and calibration curves. Cross-validation procedures were conducted. The distribution of length of stay was heavily right-skewed with a mean of 3.5 days (95% confidence interval 0.3–19.1). Provider orders were predictive of length of stay in real-time accurately forecasting discharge within a 12-hr window: 46% for patients within 1 day of discharge, 34% for patients within 2 days of discharge, and 27% for patients within 3 days of discharge. The forecast model incorporating predictive orders demonstrated significant improvements in accuracy compared with forecasts based solely on empirical and temporal information. Seventeen predictive orders were found, grouped by medication, ventilation, laboratory, diet, activity, foreign body, and extracorporeal membrane oxygenation.
Conclusions:
Provider orders reflect dynamic changes in patients’ conditions, making them useful for real-time length of stay prediction and patient flow management. Patients’ length of stay represent a major source of variability in intensive care unit resource utilization and if accurately predicted and communicated, may lead to proactive bed management with more efficient patient flow.
