Machine learning models to predict and benchmark PICU length of stay with application to children with critical bronchiolitis

Objective: To create models for prediction and benchmarking of pediatric intensive care unit (PICU) length of stay (LOS) for patients with critical bronchiolitis.

Hypothesis: We hypothesize that machine learning models applied to an administrative database will be able to accurately predict and benchmark the PICU LOS for critical bronchiolitis.

Design: Retrospective cohort study.

Patients: All patients less than 24-month-old admitted to the PICU with a diagnosis of bronchiolitis in the Pediatric Health Information Systems (PHIS) Database from 2016 to 2019.

Methodology: Two random forest models were developed to predict the PICU LOS. Model 1 was developed for benchmarking using all data available in the PHIS database for the hospitalization. Model 2 was developed for prediction using only data available on hospital admission. Models were evaluated using R² values, mean standard error (MSE), and the observed to expected ratio (O/E), which is the total observed LOS divided by the total predicted LOS from the model.

Results: The models were trained on 13,838 patients admitted from 2016 to 2018 and validated on 5254 patients admitted in 2019. While Model 1 had superior R² (0.51 vs. 0.10) and (MSE) (0.21 vs. 0.37) values compared to Model 2, the O/E ratios were similar (1.18 vs. 1.20). Institutional median O/E (LOS) ratio was 1.01 (IQR 0.90-1.09) with wide variability present between institutions.

Conclusions: Machine learning models developed using an administrative database were able to predict and benchmark the length of PICU stay for patients with critical bronchiolitis.