Predicting emergency mortality risk in traumatic brain injury: comparative analysis of machine learning and large language model GPT-5.

BACKGROUND

Artificial intelligence (AI) has become increasingly important in predicting outcomes of traumatic brain injury (TBI). Traditional machine learning (ML) models, such as support vector machines (SVMs), have shown high accuracy, whereas the potential of large language models (LLMs) in structured clinical prediction remains underexplored.

PURPOSE: This study compared the predictive performance of ML and LLM approaches (ChatGPT-5 with Thinking response mode) using the same TBI dataset and evaluated the impact of prompting strategies and threshold calibration on model reliability and clinical applicability.

METHODS

A dataset of 5,475 TBI cases with 12 clinical features was used to build an SVM model and four LLM strategies: zero-shot GPT, few-shot GPT, few-shot + chain-of-thought (CoT) GPT, and CoT-only GPT. Performance was evaluated by accuracy, sensitivity, specificity, and ROC-AUC under fixed and balanced thresholds.

RESULTS

The SVM model achieved the best overall performance (AUC = 0.920). After threshold adjustment, all LLMs reached comparable AUCs (0.902-0.919). Few-shot GPT most closely matched SVM, CoT + few-shot achieved highest sensitivity, and CoT-only favored specificity.

CONCLUSION

Proper threshold calibration enables LLMs to approximate ML accuracy while offering rapid deployment and interpretability. Prompt engineering combined with adaptive cut-off tuning may enhance clinical usability of LLM-based prediction systems.