The implementation of an Extreme Gradient Boosting machine-learning model to predict postoperative length of stay may have improved hospital bed utilization at a tertiary pediatric center, according to findings from a preimplementation and postimplementation cohort study published in JAMA Pediatrics. The model achieved 86% accuracy with a 1-night leniency. Following integration into elective surgery scheduling, the percentage of weekdays with underused capacity decreased from 33% to 10%, while the median number of elective surgical procedures increased by five per weekday.

In the study, conducted at Boston Children's Hospital, investigators analyzed 21,352 elective surgical cases for Extreme Gradient Boosting (XGBoost) model derivation and 12,522 cases for the pretest-posttest analysis. The mean age of patients was 10 years, and about 51% were female. The investigators derived the length of stay (LOS) prediction model using data from January 2018 through March 2022 and subsequently implemented it for utilization management from July 2023 through April 2024, comparing outcomes against a preimplementation period from July 2022 through April 2023.

The XGBoost model incorporated 1,367 predictors encompassing patient demographics, comorbidities classified using a pediatric adaptation of the Agency for Healthcare Research and Quality Chronic Condition Indicator system, and surgical procedure types identified via Current Procedural Terminology codes. The model was calibrated on LOS categories by number of nights (1, 2, 3, 4, 5, and 6 or more nights), with hyperparameter tuning performed using fivefold cross-validation.

Compared with alternative approaches—random forest (nearly 86% accuracy with 1-night leniency), logistic regression (85%), and k-nearest neighbor (84%)—XGBoost demonstrated superior overall performance. All of the models exhibited higher rates of LOS underprediction (20% to 25%) compared with rates of overprediction (10% to 12%).

The operational implementation employed a prospective calendar system displaying estimated postoperative census for upcoming weeks. Days projected to exceed capacity were flagged, enabling surgical schedulers to redistribute cases to lower-census days. This approach targeted census smoothing across weekdays; the model exposed patterns of low census on Mondays and Tuesdays with higher occupancy midweek.

Variation in postoperative bedded days decreased substantially following implementation. The magnitude of variation in bedded days showed the greatest reductions midweek, with reductions of 43% and 44% in the interquartile range on Wednesdays and Thursdays, respectively. Mondays, Saturdays, and Sundays demonstrated reductions ranging from 31% to 38%.

While the rates of overcapacity days increased for Tuesdays, Thursdays, and Fridays, these changes were not statistically significant. The study authors noted: "The hospital did not reach complete, full capacity on these days and no elective surgical procedures were canceled due to high bed census."

The findings addressed a persistent operational challenge in pediatric hospitals. As the study authors observed: "Historically, inpatient bed capacity has limited the volume of elective surgical cases that can be accommodated each day in a hospital." Traditional volume-capping strategies, while preventing overbooking, don't incorporate postoperative LOS estimates that could identify and mitigate both low- and high-occupancy days.

The study had several limitations. As a single-center investigation at a high-volume pediatric surgical center, generalizability may have been limited. The model was calibrated on nights rather than hours, potentially constraining granularity for capacity management. Additionally, administrative diagnostic codes—with known variation in use—informed chronic condition inputs.

"By harnessing the power of advanced analytics and predictive algorithms, health care organizations can navigate the complexities of capacity planning with greater precision and foresight, ultimately advancing their mission to deliver accessible, efficient, and high-quality care to all patients," concluded lead study author Jay G. Berry, MD, MPH, of the Division of General Pediatrics at the Boston Children's Hospital, and colleagues.

Disclosures

Ben Y. Reis, PhD, reported receiving equity in Gnomica Biotherapy and Tedence and consulting fees from Somite Therapeutics. No other disclosures were reported.

Source: JAMA Pediatrics