Healthcare-Based Multimodal Recovery Prediction for the Warfighter: A Retrospective Study

Area of Science:

• Surgical outcomes research
• Machine learning in healthcare
• Clinical informatics

Background:

• Military medics and clinicians need to understand warfighter recovery post-surgery for operational planning and resource allocation.
• Predicting short-term recovery outcomes is crucial for readiness and effective management in military treatment facilities and civilian hospitals.

Purpose of the Study:

• Develop a machine learning model to predict short-term surgical recovery outcomes.
• Identify key clinical factors influencing patient recovery post-surgery.
• Provide insights for improving military readiness and resource allocation.

Main Methods:

• Utilized the INSPIRE dataset (over 131,000 surgical procedures) from South Korea.
• Defined recovery outcome as time to discharge (≤24 hours, >24 hours, or ICU admission).
• Employed logistic regression, random forest, and XGBoost models, optimizing with 10-fold cross-validation and evaluating with custom AUC metrics. SHapley Additive Explanations identified key predictive features.

Main Results:

• An XGBoost classifier achieved high performance with a micro-AUC of 0.948 and macro-AUC of 0.815.
• Intraoperative vital signs and the department of procedure were identified as the most significant predictors of recovery.
• The model accurately predicted recovery outcomes across three defined categories.

Conclusions:

• An accurate machine learning model for predicting surgical recovery time was successfully developed.
• Identifying key predictive factors like intraoperative vital signs offers critical information for clinicians and medics.
• Future research aims to validate these findings in a US military patient cohort.