Using Interpretable Artificial Intelligence Algorithms in the Management of Blunt Splenic Trauma: Applications of Optimal Policy Trees as a Treatment Prescription Aid to Improve Patient Mortality

Abstract

Background: The identification of the optimal management for blunt splenic trauma—angioembolization (AE), splenectomy, or observation—remains a challenge. This study applies Optimal Policy Trees (OPT), an artificial intelligence (AI) model, to prescribe appropriate management and improve in-hospital mortality. Methods: OPTs were trained on patients with blunt splenic injuries in the ACS-TQIP 2013–2019 to prescribe one of the three interventions: splenectomy, angioembolization (AE), or observation. Prescriptive trees were derived in two separate patient cohorts: those who presented with a systolic blood pressure (SBP) < 70 mmHg and those with an SBP ≥ 70 mmHg. Splenic injury severity was graded using the American Association of Surgical Trauma (AAST) grading scale. Counterfactual estimation was used to predict the effects of interventions on overall in-hospital mortality. Results: Among 54,345 patients, 3.1% underwent splenic AE, 13.1% splenectomy, and 83.8% were managed with observation. In patients with SBP < 70 mmHg, AE was recommended for shock index (SI) < 1.5 or without transfusion, while splenectomy was indicated for SI ≥ 1.5 with transfusion. For patients with SBP ≥ 70 mmHg, AE was recommended for AAST grades 4–5, or grades 1–3 with SI ≥ 1.2; observation was recommended for grades 1–3 with SI < 1.2. Predicted mortality using OPT-prescribed treatments was 18.4% for SBP < 70 mmHg and 4.97% for SBP ≥ 70 mmHg, compared to observed rates of 36.46% and 7.60%, respectively. Conclusions: Interpretable AI models may serve as a decision aid to improve mortality in patients presenting with a blunt splenic injury. Our data-driven prescriptive OPT models may aid in prescribing the appropriate management in this patient cohort based on their characteristics.