Forecasting Pediatric Trauma Volumes: Insights From a Retrospective Study Using Machine Learning

Abstract

Introduction

Rising pediatric firearm-related fatalities in the United States strain Trauma Centers. Predicting trauma volume could improve resource management and preparedness, particularly if daily forecasts are achievable. The aim of the study is to evaluate various machine learning models’ accuracy on monthly, weekly, and daily data.

Methods

The retrospective study utilized trauma data between June 1, 2013, and October 31, 2023, from a level I/II pediatric trauma center. Data were organized monthly, weekly, and daily, which further delineated into seven groups, yielding 21 cohorts. Models were evaluated using time-series forecasting metrics. In addition, the models were tested for real-world applicability by forecasting trauma volumes 3 mo, 12 wk, and 31 d ahead for monthly, weekly, and daily predictions respectively. The predicted values were then compared with the actual data.

Results

The total of 12,144 patients’ data was utilized to create and evaluate models. 14 forecasting models for each of 21 groups were developed. Monthly predictions generally outperformed weekly and daily ones. Although the Silverkite model excelled in monthly predictions, the one-dimensional convolutional layer model was most accurate for daily predictions. Real-life simulations showed the Prophet model performing best for monthly predictions, with no clear winner for weekly predictions.

Conclusions

This study found monthly forecasting most accurate. Although many models outperformed their Naïve counterparts, performance varied by grouping. Real-world simulations confirmed these findings. Despite high accuracy in monthly predictions, the study's generalizability is limited, and daily trauma prediction needs improvement.