QJM. 2025 Aug 8:hcaf180. doi: 10.1093/qjmed/hcaf180. Online ahead of print.
ABSTRACT
BACKGROUND: Kawasaki disease shock syndrome (KDSS), a severe and uncommon phenomenon, lacks effective predictive models for early identification.
AIM: This study aimed to establish a new predictive model for KDSS using machine learning.
DESIGN: Single-center, retrospective analysis.
METHODS: Data of 746 children with KD admitted between July 2021 and June 2023 were collected including demographics, laboratory test results before intravenous immunoglobulin, and echocardiography results. Data were divided into training and testing sets in a 7:3 ratio. After feature engineering, predictive models were built using random forest (RF), logistic regression (LR), and Light Gradient Boosting Machine (LightGBM). Model performance was evaluated using area under the receiver operating characteristic curve (AUC), confusion matrix, average accuracy from five-fold cross-validation, while also analyzing misclassified cases. A simple early prediction tool was created based on the optimal model. Prospective data from five KDSS patients admitted between January and June 2024 and that of 15 randomly selected non-shock KD patients were used for external validation.
RESULTS: CD3+ lymphocyte percentage(CD3+%) had the greatest impact on the model and was an important predictive factor for KDSS, followed by neutrophil-to-lymphocyte(NLR) ratio and Interleukin-6(IL-6). The LightGBM model performed best (AUC, 0.9388; average accuracy, 0.9675; 95% CI, 0.9612, 0.9737). Nine patients were misclassified (4.02%). RF and LR models showed slightly lower performance than the LightGBM model (prospective validation AUC, 0.9000; accuracy, 0.8500).
CONCLUSION: We constructed an early prediction model for KDSS and performed preliminary validation. This web-based prediction tool may assist clinicians in identifying high-risk pediatric patients to enhance monitoring/treatment.
PMID:40795391 | DOI:10.1093/qjmed/hcaf180
