Field Foresight—Predicting the Need for Massive Transfusion, ICU Utilization, and Mechanical Ventilation for Traumatically Injured Patients

Military Medicine, vol. 191, pp. e1971–e1979

Abstract

INTRODUCTION: Expeditiously predicting outcomes is essential to allocating blood and intensive care resources. We hypothesize the use of external injuries and vital signs collected early after injury will improve artificial intelligence (AI) triage performance of civilian and military patients, compared to benchmark algorithms using only vital signs. To test this, we developed the Field AI Triage (FAIT) tool. MATERIALS AND METHODS: We leveraged civilian (American College of Surgeons Trauma Quality Improvement Program, TQIP) and military (Department of Defense Trauma Registry, DoDTR) datasets to develop a logistic regression model to guide triage of blunt and penetrating traumatic injuries. Inputs: vital signs and contextual features, including descriptions of injury mechanism and location. Outputs: need for massive transfusion, mechanical ventilation, and ICU utilization. The model was tested with 10-fold cross validation. The primary performance metric was receiver operating characteristic area under the curve (ROC AUC). RESULTS: Data from 786,957 civilian patients (TQIP) and 8,946 military patients (DoDTR) were used to develop FAIT. For civilian patients, FAIT improved AUC over the vitals-only benchmark: 0.89 ± 0.00 vs. 0.70 ± 0.01 for massive transfusion (401,337 patient subset), 0.90 ± 0.00 vs. 0.65 ± 0.00 for mechanical ventilation, and 0.78 ± 0.00 vs. 0.61 ± 0.01 for ICU admittance. For military patients, performance improvements were of 0.88 ± 0.02 vs. 0.67 ± 0.04 for massive transfusion, 0.90 ± 0.01 vs. 0.76 ± 0.02 for mechanical ventilation, and 0.79 ± 0.01 vs. 0.69 ± 0.02 for ICU admittance. Sensitivity analysis provides insight on feature importance. CONCLUSIONS: FAIT supports assessment at the point of injury and demonstrates significant capability in forecasting resource allocation for civilian and military trauma patients across a wide variety of mechanisms.

Authors 10

1. Allan E. Stolarski corresponding

   Boston University · Boston Medical Center · University of Pennsylvania

   Affiliation as printed

   Division of Trauma and Acute Care Surgery, Boston Medical Center, Boston University School of Medicine , Boston, MA 02118,

   Division of Traumatology, Surgical Critical Care, University of Pennsylvania , Philadelphia, PA 19104,

2. Kevin J. Brady

   MIT Lincoln Laboratory

   Affiliation as printed

   MIT Lincoln Laboratory , Lexington, MA 02421,

3. Jonathan Daniel Stallings
   Affiliation as printed

   Joint Trauma System , Fort Sam Houston, TX 78234,

4. Dane R. Scantling

   Boston University · Boston Medical Center

   Affiliation as printed

   Division of Trauma and Acute Care Surgery, Boston Medical Center, Boston University School of Medicine , Boston, MA 02118,

5. Crisanto M. Torres

   Boston University · Boston Medical Center

   Affiliation as printed

   Division of Trauma and Acute Care Surgery, Boston Medical Center, Boston University School of Medicine , Boston, MA 02118,

6. Ava K. Mokhtari

   The University of Texas Health Science Center at Houston

   Affiliation as printed

   Department of Surgery, The University of Texas Health Science Center and the McGovern School of Medicine , Houston, TX 77030,

7. Charlie J. Nederpelt Aachen

   Leiden University Medical Center

   Affiliation as printed

   Leiden University Medical Center , Leiden 2300 RC,

8. Ryan J. McKindles

   MIT Lincoln Laboratory

   Affiliation as printed

   MIT Lincoln Laboratory , Lexington, MA 02421,

9. Brian A. Telfer

   MIT Lincoln Laboratory

   Affiliation as printed

   MIT Lincoln Laboratory , Lexington, MA 02421,

10. Noelle Nugent Saillant

    Boston University · Boston Medical Center

    Affiliation as printed

    Division of Trauma and Acute Care Surgery, Boston Medical Center, Boston University School of Medicine , Boston, MA 02118,