A Pilot Machine Learning Study Using Trauma Admission Data to Identify Risk for High Length of Stay.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Trauma Surgery

Background:

• Trauma patient resource needs vary significantly based on injury type and mechanism.
• Predicting prolonged hospital length of stay (LOS) is crucial for resource allocation and patient management.
• Current prediction methods often rely on data not available at admission.

Purpose of the Study:

• To develop a predictive tool for prolonged LOS using only data available at the time of trauma patient admission.
• To leverage machine learning, specifically artificial neural networks, for enhanced predictive accuracy.
• To optimize the model for identifying patients in the highest quartile of LOS.

Main Methods:

• Utilized data from an urban level one adult trauma center registry (1/1/2014-3/31/2019).
• Included trauma patients with LOS > 1 day, excluding those with shorter stays.
• Trained single-layer and deep artificial neural networks, optimizing for Area Under the Receiver Operator Characteristic Curve (AUROC).

Main Results:

• The study included 2953 trauma patients with LOS > 1 day.
• Deep neural network achieved an AUROC of 0.80 (95% CI: 0.786-0.814).
• Model demonstrated high specificity (0.95) and overall accuracy (0.79) in predicting prolonged LOS.

Conclusions:

• Machine learning models can effectively predict prolonged LOS in trauma patients using only admission data.
• The developed tool offers high specificity, identifying patients likely to require extended care.
• Early identification facilitates proactive resource allocation and disposition planning for trauma patients.