Enhancing the Management of Deployment-Limiting Medical Conditions in the U.S. Department of Navy Leveraging Electronic Health Records and Predictive Analytics Methods.

INTRODUCTION

Deployment-limiting medical conditions (DLMCs) such as debilitating injuries and conditions may interfere with the ability of military service members (SMs) to deploy. SMs in the United States (U.S.) Department of the Navy (DoN) with DLMCs who are not deployable should be placed in the medically restricted status of limited duty (LIMDU) or referred to the Physical Evaluation Board (PEB) for Service retention determination. It is critical to identify SMs correctly and promptly with DLMCs and predict their return-to-duty (RTD) to ensure the combat readiness of the U.S. Military. In this study, an algorithmic approach was developed to identify DoN SMs with previously unidentified DLMCs and predict whether SMs on LIMDU will be able to RTD.

MATERIALS AND METHODS

Five years of historical data (2016-2022) were obtained from inpatient and outpatient datasets across direct and purchased care from the Military Health System (MHS) Data Repository (MDR). Key fields included International Classification of Diseases diagnosis and procedure codes, Current Procedure Terminology codes, prescription medications, and demographics information such as age, rank, gender, and service. The data consisted of 44,580,668 medical encounters across 1,065,224 SMs. To identify SMs with unidentified DLMCs, we developed an ensemble model combining outputs from multiple machine learning (ML) algorithms. When the ML ensemble model predicted a SM to have high risk scores, despite appearing healthy on administrative reports, their case was reviewed by expert clinicians to investigate for previously unidentified DLMCs; and such feedback served to validate the developed algorithms. In addition, leveraging 1,735,422 encounters (60,433 SMs) from LIMDU periods, we developed four separate ML models to estimate RTD probabilities for SMs after each medical encounter and predict the final LIMDU outcome.

RESULTS

The ensemble model had 0.91 area under the receiver operating characteristic curve (AUROC). Out of 236 (round one) and 314 (round two) SMs reviewed by clinicians, 127 (54%) and 208 (66%) SMs were identified with a previously unidentified or undocumented DLMC, respectively. Regarding predicting RTD for SMs placed on LIMDU, the best performing ML model achieved 0.76 AUROC, 68% sensitivity, and 71% specificity.

CONCLUSION

Our research highlighted potential benefits of using predictive analytics in a medical assessment to identify SMs with DLMCs and to predict RTD outcomes once placed on LIMDU. This capability is being deployed for real-time clinical decision support to enhance health care provider's deployability assessment capability, improve accuracy of the DLMC population, and enhance combat readiness of the U.S Military.