Scholarlink

<p>The American Society of Anesthesiologists Physical Status classification suffers from substantial inter-rater variability, compromising perioperative risk assessment. Large language models have demonstrated remarkable capabilities in medical text processing, but their application to anesthetic risk prediction remains unexplored. This study compared large language model performance against traditional machine learning approaches for automated ASA classification. This retrospective study analyzed 21,049 surgical records from the MOVER database (2017–2022), utilizing structured preoperative data including patient demographics, laboratory values, medication records, and procedural information for ASA I-IV classifications. We compared five large language models (GPT-4, GPT-4 Turbo, GPT-4o, Gemini 1.5 Flash, Gemini 1.5 Pro) against eight traditional machine learning algorithms using 10-fold cross-validation. Four training approaches were implemented: zero-shot learning, few-shot learning, few-shot with machine learning features, and few-shot with large language model features. Performance was evaluated using precision, recall, F₁-score, AUROC, and AUPRC. GPT-4o Few Shot Feature LLM achieved superior performance with 66% accuracy, F₁-score of 0.65, and AUROC of 0.79, substantially outperforming the best traditional method LightGBM (54% accuracy), representing a 12%-point improvement in classification accuracy. Large language models demonstrated excellent performance in intermediate risk categories (ASA II-III) with precision 0.60–0.75 and recall 0.60–0.78, covering 85% of surgical patients. However, all models showed low recall rates for ASA IV patients (28%), indicating challenges identifying highest-risk patients. Feature importance analysis revealed convergent validity between approaches, consistently identifying weight, age, and medication complexity as primary predictors. Large language models outperformed traditional machine learning for ASA classification in this dataset, particularly in intermediate-risk patients comprising most surgical cases. Their superior ability to interpret structured clinical text, including procedure names, medication records, and laboratory results, demonstrates potential for reducing inter-rater variability. However, low recall for high-risk patients necessitates human oversight and thus positions these models as decision support tools. Multi-institutional validation is essential before clinical implementation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparative Analysis of Large Language Models and Machine Learning for ASA Classification Using Structured Electronic Health Record Data

  • 期刊论文

  • 发表日期：

    2026年4月29日

查看全文
  • Shih-Yu Ko,
  • Ting-Yun Huang,
  • Ming-Siang Huang,
  • Yi-Hsuan Lin,
  • Yung-Cheng Su,
  • Yung-Chun Chang

摘要

The American Society of Anesthesiologists Physical Status classification suffers from substantial inter-rater variability, compromising perioperative risk assessment. Large language models have demonstrated remarkable capabilities in medical text processing, but their application to anesthetic risk prediction remains unexplored. This study compared large language model performance against traditional machine learning approaches for automated ASA classification. This retrospective study analyzed 21,049 surgical records from the MOVER database (2017–2022), utilizing structured preoperative data including patient demographics, laboratory values, medication records, and procedural information for ASA I-IV classifications. We compared five large language models (GPT-4, GPT-4 Turbo, GPT-4o, Gemini 1.5 Flash, Gemini 1.5 Pro) against eight traditional machine learning algorithms using 10-fold cross-validation. Four training approaches were implemented: zero-shot learning, few-shot learning, few-shot with machine learning features, and few-shot with large language model features. Performance was evaluated using precision, recall, F₁-score, AUROC, and AUPRC. GPT-4o Few Shot Feature LLM achieved superior performance with 66% accuracy, F₁-score of 0.65, and AUROC of 0.79, substantially outperforming the best traditional method LightGBM (54% accuracy), representing a 12%-point improvement in classification accuracy. Large language models demonstrated excellent performance in intermediate risk categories (ASA II-III) with precision 0.60–0.75 and recall 0.60–0.78, covering 85% of surgical patients. However, all models showed low recall rates for ASA IV patients (28%), indicating challenges identifying highest-risk patients. Feature importance analysis revealed convergent validity between approaches, consistently identifying weight, age, and medication complexity as primary predictors. Large language models outperformed traditional machine learning for ASA classification in this dataset, particularly in intermediate-risk patients comprising most surgical cases. Their superior ability to interpret structured clinical text, including procedure names, medication records, and laboratory results, demonstrates potential for reducing inter-rater variability. However, low recall for high-risk patients necessitates human oversight and thus positions these models as decision support tools. Multi-institutional validation is essential before clinical implementation.