Scholarlink

<p>Postoperative systemic inflammatory response syndrome (SIRS) and sepsis following ureteroscopic lithotripsy (URS) are uncommon but potentially life-threatening complications. No prior study has evaluated large language models (LLMs) for predicting post-URS SIRS/sepsis using exclusively preoperative parameters. This study evaluated the discriminative performance of LLMs in this setting. This retrospective case–control study included 210 patients who underwent URS between January 2024 and December 2025 (42 SIRS/sepsis cases, 168 controls). Risk scores (0–100) were generated using GPT-4o-mini and Gemini 2.5 Flash with guideline-based and intuitive prompting strategies. Discriminative performance was assessed using receiver operating characteristic analysis and area under the curve (AUC). Model comparisons were performed using the DeLong test. Patients who developed SIRS/sepsis were older and more frequently female, with larger stones, higher hydronephrosis grade, lower stone density, increased inflammatory markers (urine WBC, leukocyte esterase, neutrophil-to-lymphocyte ratio), higher rates of positive urine culture, and more frequent radiological inflammatory findings (all p&lt;0.001). Total WBC did not differ between groups (p=0.698). All model–prompt combinations showed high discriminative performance (AUC 0.952–0.973). The highest performance was achieved by the GPT-4o-mini intuitive model (AUC 0.973; sensitivity 92.9%; specificity 95.2%). No significant differences were observed between models or prompting strategies. Performance remained high after excluding patients with positive urine culture or preoperative antibiotic use (AUC 0.896–0.943). LLMs demonstrated high discriminative performance for predicting postoperative SIRS/sepsis using only preoperative data, suggesting potential utility as accessible clinical decision-support tools for perioperative risk stratification in URS. Prospective multicenter validation is required 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

Preoperative prediction of postoperative SIRS and sepsis after ureteroscopic lithotripsy: discriminative performance of large language models

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Hüseyin Koçakgöl,
  • Deniz Öztürk Koçakgöl

摘要

Postoperative systemic inflammatory response syndrome (SIRS) and sepsis following ureteroscopic lithotripsy (URS) are uncommon but potentially life-threatening complications. No prior study has evaluated large language models (LLMs) for predicting post-URS SIRS/sepsis using exclusively preoperative parameters. This study evaluated the discriminative performance of LLMs in this setting. This retrospective case–control study included 210 patients who underwent URS between January 2024 and December 2025 (42 SIRS/sepsis cases, 168 controls). Risk scores (0–100) were generated using GPT-4o-mini and Gemini 2.5 Flash with guideline-based and intuitive prompting strategies. Discriminative performance was assessed using receiver operating characteristic analysis and area under the curve (AUC). Model comparisons were performed using the DeLong test. Patients who developed SIRS/sepsis were older and more frequently female, with larger stones, higher hydronephrosis grade, lower stone density, increased inflammatory markers (urine WBC, leukocyte esterase, neutrophil-to-lymphocyte ratio), higher rates of positive urine culture, and more frequent radiological inflammatory findings (all p<0.001). Total WBC did not differ between groups (p=0.698). All model–prompt combinations showed high discriminative performance (AUC 0.952–0.973). The highest performance was achieved by the GPT-4o-mini intuitive model (AUC 0.973; sensitivity 92.9%; specificity 95.2%). No significant differences were observed between models or prompting strategies. Performance remained high after excluding patients with positive urine culture or preoperative antibiotic use (AUC 0.896–0.943). LLMs demonstrated high discriminative performance for predicting postoperative SIRS/sepsis using only preoperative data, suggesting potential utility as accessible clinical decision-support tools for perioperative risk stratification in URS. Prospective multicenter validation is required before clinical implementation.