Artificial Intelligence Length-of-Stay Forecasting and Pediatric Surgical Capacity

Comprehensive Summary

In this study, Berry et al. analyze how artificial intelligence (AI) can be used to efficiently and accurately estimate postoperative length of stay (LOS) at a major freestanding pediatric hospital in the US, aiming to reduce underuse of hospital beds while optimizing the number of surgical procedures performed. First, Berry et al. trained and tested 4 AI models (XGBoost, random forest, logistic regression, and k-nearest neighbor) to predict LOS based on data from 21,352 elective surgical cases from 2018-2022. The predictive data included demographic characteristics, patient comorbidities, surgical procedures, and median and 90th-percentile LOS for those procedures. Then, a prospective evaluation of the best model was conducted using 12,522 cases over 2022-2024. In testing, XGBoost achieved 67.7% accuracy in predicting LOS exactly and reached 85.6% within a one-night margin of error, marginally better than the other models. The mean absolute error was 0.6 days, indicating that incorrect predictions remained very close to the true LOS. XGBoost was then implemented by the hospital to optimize hospital bed use and surgical procedure performance. The median number of elective surgical procedures increased by 5 for each weekday, underuse of beds decreased significantly from 33% to 10% (p<0.001), and there was no significant increase in days with excessive capacity. Overall, this model effectively predicted LOS, reduced bed underuse, and improved access to surgery, without overburdening the hospital or increasing cancellations. This study’s focus on implementing this model in a large pediatric hospital also highlighted its potential to significantly benefit clinical practices and hospital systems. The authors note that while it was not statistically significant, there was an increase in the percentage of overcapacity days which can negatively impact patient outcomes. Further steps include incorporating this model into other hospitals and surgical specialties, and adding more socioeconomic features to its design.

Outcomes and Implications

This study by Berry et al. is relevant to AI, pediatric surgery, and medicine because it demonstrates the longitudinal success of AI in clinical practice. Rather than serving solely as a predictive tool, the model in this paper was used to directly influence elective surgery scheduling. Especially since pediatric surgery involves a highly heterogeneous patient population with a wide range of recovery times, the success of this model is particularly useful for ensuring timely surgical care. Future implementation of this technology in other sections and hospitals could similarly improve workflow and management.