Scholarlink

Background <p>Intravenous immunoglobulin (IVIG) is the standard initial treatment for Kawasaki disease (KD). However, approximately 20% of patients are resistant, who have a significantly increased risk of coronary artery lesions. Early prediction of IVIG resistance is crucial for improving outcomes, but traditional clinical scoring systems exhibit limited generalizability and limited predictive performance.</p> Objective <p>To systematically evaluate the accuracy of machine learning (ML) models in predicting IVIG-resistant KD through meta-analysis, and provide evidence-based support for clinical application.</p> Methods <p>A comprehensive search of PubMed, Cochrane Library, and Embase was performed from inception to September 15, 2025. The novel PROBAST (Prediction Model Risk of Bias Assessment Tool) was used to assess the risk of bias of the ML models. A random-effects model was used to synthesize the area under the curve (AUC), and a bivariate mixed-effects model was applied for sensitivity and specificity analysis. Subgroup analysis was conducted based on model type and definition criteria.​</p> Results <p>A total of 22 studies (32,678 patients) were included. The pooled AUC was 0.80 (95% CI: 0.75–0.84) for training sets and 0.80 (95% CI: 0.76–0.84) for internal validation sets. The pooled AUC for external validation sets was 0.74 (95% CI: 0.68–0.79). Subgroup analysis revealed that complex tree-based ensemble models demonstrated significantly superior discriminative ability (AUC = 0.85, 95% CI: 0.80–0.90). After excluding one outlier study, the pooled sensitivity and specificity for internal validation sets were 0.79 (95% CI: 0.68–0.87) and 0.85 (95% CI: 0.68–0.94), respectively. Key predictors frequently included in ML models were C-reactive protein, platelet count, neutrophil percentage and hemoglobin.</p> Conclusions <p>Our research demonstrates that ML models exhibit favorable and promising performance in predicting IVIG-resistant KD, outperforming traditional scoring systems, which provides support for the future development of ML in this field. Notwithstanding, cautious interpretation is required given the moderate external validation efficacy and high risk of bias in most included studies.</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

Machine learning prediction models for intravenous immunoglobulin resistance in Kawasaki disease: a meta-analysis

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月18日

查看全文
  • Chuankun Zhang,
  • Yongle Zhao,
  • Yangzhongyi Liu,
  • Chang Liu,
  • Yi Wei,
  • Nuo Xu,
  • Hengpeng Luo,
  • Penghui Yang,
  • Qijian Yi

摘要

Background

Intravenous immunoglobulin (IVIG) is the standard initial treatment for Kawasaki disease (KD). However, approximately 20% of patients are resistant, who have a significantly increased risk of coronary artery lesions. Early prediction of IVIG resistance is crucial for improving outcomes, but traditional clinical scoring systems exhibit limited generalizability and limited predictive performance.

Objective

To systematically evaluate the accuracy of machine learning (ML) models in predicting IVIG-resistant KD through meta-analysis, and provide evidence-based support for clinical application.

Methods

A comprehensive search of PubMed, Cochrane Library, and Embase was performed from inception to September 15, 2025. The novel PROBAST (Prediction Model Risk of Bias Assessment Tool) was used to assess the risk of bias of the ML models. A random-effects model was used to synthesize the area under the curve (AUC), and a bivariate mixed-effects model was applied for sensitivity and specificity analysis. Subgroup analysis was conducted based on model type and definition criteria.​

Results

A total of 22 studies (32,678 patients) were included. The pooled AUC was 0.80 (95% CI: 0.75–0.84) for training sets and 0.80 (95% CI: 0.76–0.84) for internal validation sets. The pooled AUC for external validation sets was 0.74 (95% CI: 0.68–0.79). Subgroup analysis revealed that complex tree-based ensemble models demonstrated significantly superior discriminative ability (AUC = 0.85, 95% CI: 0.80–0.90). After excluding one outlier study, the pooled sensitivity and specificity for internal validation sets were 0.79 (95% CI: 0.68–0.87) and 0.85 (95% CI: 0.68–0.94), respectively. Key predictors frequently included in ML models were C-reactive protein, platelet count, neutrophil percentage and hemoglobin.

Conclusions

Our research demonstrates that ML models exhibit favorable and promising performance in predicting IVIG-resistant KD, outperforming traditional scoring systems, which provides support for the future development of ML in this field. Notwithstanding, cautious interpretation is required given the moderate external validation efficacy and high risk of bias in most included studies.