Evaluation and analysis of clinical outcome prediction for trauma patients based on machine learning

Abstract

PURPOSE: The study was to compare the predictive performance of multiple models for clinical outcomes in trauma patients. To provide decision-making support for the clinical management of trauma patients by analyzing and ranking the importance of predictive indicators. METHODS: The data were collected from the sub-system of the national trauma union database (2022 - 2024) in Shanghai East Hospital. The terminal outcomes (death or survive) were the final predictive variables. Lasso regression was employed to screen factors affecting clinical trauma treatment pathways, which were then used as indicators for the final model construction. Logistic regression, random forest, support vector machine, XGBoost, and neural network were used for model evaluation. The predictive performance of each model was assessed through 5 evaluation index including area under curve, accuracy, recall, precision, and F1 score. SHapley Additive exPlanations (SHAP) analysis was employed to assess the importance of variables and visualize the contribution and impact of variables on trauma patient outcomes. RESULTS: Lasso regression selected a total of 9 variables for the final model construction. The variables were: Trauma index score, diastolic blood pressure, body temperature, blood oxygen saturation, venous access, injury severity score (ISS), abbreviated injury scale (AIS) score, residence time in emergency department, and number of surgeries. XGBoost had the highest sensitivity (0.800) and Youden's index (0.775), demonstrating strong recognition ability for positive samples. Logistic regression achieved the best accuracy (0.985) and Kappa coefficient (0.598), with stronger overall classification consistency. Neural network had the lowest sensitivity (0.450) and Youden's index (0.441), and its recognition ability for positive samples was relatively weak. All models had relatively high specificity (0.976 - 0.991), indicating generally good recognition ability for negative samples, but the differences in Precision and F1 score reflected the performance divergence of different models in "reliability of predicting positives" and "precision-recall balance". SHAP analysis showed that the feature "venous access" occupied the most important position among all features, with a mean absolute SHAP value of about 0.012, which had the most significant impact on the model's prediction and may be the key factor in the model's decision-making. In addition, the AIS score had a mean absolute SHAP value of about 0.011, and the ISS was about 0.008, both of which significantly affected the model's prediction. Moreover, other features such as "truama index Score" and "residence time in emergency room" also played important roles in the model's prediction, although their impacts were relatively smaller. CONCLUSION: We analyzed the optimal model from 5 aspects: (1) receiver operating characteristic curve, (2) 5 indicators, (3) the trade-off between precision and recall, (4) scenario application: emergency settings, and (5) economic benefits. XGBoost was recommended as the first choice based on the performance, but external validation was required to ensure clinical applicability. SHAP analysis enhanced model transparency, helping rescue staff to understand the degree of influence of variables and provide a reference for clinical decision-making. Furthermore, key variables (such as ISS and AIS scores) aligned with medical consensus, verifying the credibility of the model. In the future, the researches should validate the model in real-world scenarios and continuously optimize features and algorithms to enhance practicality.

Ähnliche Arbeiten