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

In Roles 3 and 4 military treatment facilities and civilian hospitals, medics, unit commanders, and clinicians require knowledge of warfighter recovery following injury or surgery to aid operational planning, improve readiness and conduct resource allocation. In this study we developed a machine learning model using pre- and intraoperative patient surgical data from electronic health records to predict short-term recovery outcomes following surgery and to derive insights into clinical factors driving recovery.

We used the INSPIRE (INformative Surgical Patient dataset for Innovative Research Environment) dataset, a publicly available, deidentified dataset from PhysioNet containing perioperative data from over 131,000 surgical procedures performed from 2011 to 2020 in South Korea to conduct our study. Data were originally collected under Institutional Review Board at Seoul National University Hospital and released publicly by their Institutional Data Review Board. Data contain patient demographics, surgical metadata, preoperative laboratory values, and pre-, post- and intraoperative vital signs. We defined recovery outcome as the time from the end of operation to discharge, which we binned into 3 categories: recovery in greater than 24 hours, recovery in 24 hours or less, or admission to the intensive care unit (ICU). We built several machine learning models (logistic regression, random forest and XGBoost) to predict our multiclass recovery outcome. Models were optimized with a 10-fold cross-validation procedure and evaluated with a custom area under the curve (AUC) metric where the predicted and true labels were one-hot encoded and a one-vs.-rest approach was used in calculation. We applied SHapley Additive Explanations to the features in the training data to determine which ones contributed most heavily to model predictions.

The best performing model was an XGBoost classifier, which reported a micro-AUC (samples given equal weight) of 0.948 and macro-AUC (classes given equal weight) of 0.815 on the 10% held out validation data. We found that intraoperative vital signs and department of procedure were the best indicators of recovery (>24 hours, ≤24 hours, or ICU).

We developed an accurate model predicting time to recovery following surgery. Knowledge of important recovery predictive factors, such as intraoperative vital signs, provide crucial information to clinicians and army medics to inform recovery and potentially return to duty and required readiness levels. In the future, we aim to replicate this work in a clinical cohort in the United States primarily focused on a military population.