(G-250) Severity Scoring for Traumatic Injury: From ED to ICU

Introduction:: Severity scores are routinely used for the classification or stratification of patients for clinical care and in the evaluation of clinical outcomes [1]. Such scores find the highest utility as a standardization instrument in the study of large cohorts of patients. The advent of pervasive tools and capabilities of Artificial Intelligence (AI) and Machine Learning (ML) augments the significance of applying such scores by enabling personalized patient care, in which case the score has the data richness of a disease biomarker (Figure 1). Recently introduced scores that capture the rich pathophysiological data of the Emergency Department (ED) and the Intensive Care Unit (ICU) serve to highlight this trend [2]. The (HISS) score seeks to stratify patients in the ED based on the fusion of diverse pathophysiological data such as glucose, lactate, pH, potassium, and oxygen tension. On the other hand, acuity assessment in the ICU seeks to dynamically classify patients based on their severity level using indicators of continuous renal replacement therapy (CRRT), ventilation, vasopressors infusion, and blood transfusion. Judiciously developed and deployed, such scores have the potential to address healthcare delivery disparities.


Materials and Methods:: One hundred static instances of Sensible Fictitious Rationalized Patient (SFRP) multi-biomarker data, glucose, lactate, pH, potassium, and oxygen tension, to serve as a patient-specific attribute in hemorrhagic trauma were synthetically generated. The Hemorrhage Intensive Severity and Survivability (HISS) score assigned by five clinically active physician experts (100[5]). The HISS score stratifies the criticality of the trauma patient as: low(0), guarded(1), elevated(2), high(3), and severe(4). Standard classifier algorithms, including linear support vector machine (SVM-L), multi-class ensemble bagged decision tree (EBDT), artificial neural network with Bayesian regularization (ANN:BR), and possibility rule-based using function approximation (PRBF), were evaluated for their potential to classify and predict a HISS score. Similarly, Catboost, a classification ML model, was evaluated for dynamic acuity assessment in the ICU every 12 hours. Using EHR from 46,929 ICU stays at UF Health Shands Hospital, 300 temporal features (vital signs, medications, laboratory test results, assessment scores) and 502 static features (patient demographics, admission information, medication and lab results history) were extracted. Calculation of 16 statistical attributes on temporal features yielded 5,302 final features. Acuity for a 12-hour window was assessed using CRRT, ventilation, vasopressor infusion, blood transfusion, discharge, and death indicators, resulting in four outcomes: stability, instability, discharge, and mortality. The dataset was split into development and evaluation sets using an 80:20 ratio. Five-fold cross validation was performed on the development set to train/validate the model, and 100-iteration bootstrap on the evaluation set to calculate 95% confidence intervals (CI) using area under the receiving operator characteristic curve (AUROC) to measure performance.


Results, Conclusions, and Discussions:: SVM-L, EBDT, ANN:BR, and PRBF generated score predictions with testing accuracies (majority vote) corresponding to 0.91±0.06, 0.93±0.04, 0.92±0.07, and 0.92±0.03, respectively, with no statistically significant difference (p >0.05). Targeted accuracies of 0.99 and 0.999 could be achieved with SFRP data size and clinical expert scores of 147[7] (0.99) and 154[9] (0.999), respectively [3]. The CatBoost model had a mean AUROC for ICU acuity assessment every 12-hour window of 0.89 (95% CI 0.89-0.89), with performance for individual acuity outcomes being 0.96 AUROC (95% CI 0.96-0.97) for mortality, 0.84 AUROC (95% CI 0.84-0.85) for discharge, 0.86 AUROC (95% CI 0.86-0.86) for stability, and 0.90 AUROC (95% CI 0.90-0.90) for instability. While not molecular biomarkers in the strictest sense, such scores are often informed by molecular indicators of pathophysiology and find utility like traditional biomarkers. The predictions of the data-driven model, in conjunction with an adjunct multi-analyte biosensor intended for point-of-care continual monitoring of trauma patients, can aid in patient stratification and triage decision-making. Similarly, the predictions of the CatBoost models can provide early warnings to clinical staff and be used for dynamic decision support in the ICU.


Acknowledgements (Optional): :

References (Optional): : [1] Havens JM, Columbus AB, Seshadri AJ, Brown CVR, Tominaga GT, Mowery NT, et al. Risk stratification tools in emergency general surgery. Trauma Surgery & Acute Care Open. 2018;3:e000160.

[2] Shickel B, Balch J, Aggas JR, Loftus TJ, Kotanen CN, Rashidi P, et al. Scoring for Hemorrhage Severity in Traumatic Injury. In: Rajendram R, Preedy VR, Patel VB, editors. Biomarkers in Trauma, Injury and Critical Care. Cham: Springer International Publishing; 2022. p. 1-18.

[3] Bhat A, Podstawczyk D, Walther BK, Aggas JR, Machado-Aranda D, Ward KR, et al. Toward a hemorrhagic trauma severity score: fusing five physiological biomarkers. Journal of Translational Medicine. 2020;18:348.