Cheon Eun Jung, Kim Gi Beom, Park Seung
Chungbuk National University Hospital, Cheongju, Korea.
Department of Pediatrics, Chungbuk National University Hospital, 776, 1 Sunhwan-ro, Seowon-gu, Cheongju-si, Chungcheongbuk-do, Republic of Korea.
Sci Rep. 2025 Jan 6;15(1):903. doi: 10.1038/s41598-025-85394-4.
Kawasaki disease (KD) is a leading cause of acquired heart disease in children, often resulting in coronary artery complications such as dilation, aneurysms, and stenosis. While intravenous immunoglobulin (IVIG) is effective in reducing immunologic inflammation, 10-15% of patients do not respond to initial therapy, and some show resistance even after two consecutive treatments. Predicting which patients will not respond to these two IVIG treatments is crucial for guiding treatment strategies and improving outcomes. This study aimed to forecast resistance to two consecutive IVIG treatments using advanced machine learning models based on clinical and laboratory data. Data from the 9th National Kawasaki Disease Patient Survey by the Korean Kawasaki Disease Society encompassing 15,378 patients (mean age 33.0 ± 24.8 months; sex ratio 1.4:1) were used. Clinical and laboratory findings included white blood cell count, absolute neutrophil count (ANC), platelet count, erythrocyte sedimentation rate, serum protein, aspartate aminotransferase, alanine aminotransferase, total bilirubin, N-terminal pro-brain natriuretic peptide, and presence of pyuria. Machine learning models, including Logistic Regression (LR), Multi-Layer Perceptron (MLP), Random Forest (RF), CATBoost, Explainable Boosting Machine (EBM), and Gradient Boosting Machine (GBM), were applied to predict treatment resistance. The machine learning models achieved Area Under the Receiver Operating Characteristic Curve (AUROC) values between 0.664 and 0.791, with the GBM model exhibiting the highest AUROC of 0.791. Analysis of feature importance revealed that ANC, serum protein, platelet count, and C-reactive protein (CRP) levels were the most significant predictors of treatment resistance. The cutoff values for these predictors were 7,860/mm³ for ANC, 7.0 g/dL for serum protein, 519,000/mm³ for platelet count, and 10.4 mg/dL for CRP. Among the patients, 12.2% were refractory to the first IVIG infusion, and 2.8% did not respond to the second IVIG treatment. Additionally, 13.1% of these patients had confirmed coronary artery dilatation (CAD) in the acute phase, and 4.7% developed CAD after the acute phase. Machine learning models effectively predict resistance to consecutive IVIG treatments, allowing for early identification of high-risk patients. Key predictors include ANC, serum protein, platelet count, and CRP levels. These findings can guide personalized treatment strategies and improve outcomes for Kawasaki Disease.
川崎病（KD）是儿童后天性心脏病的主要病因，常导致冠状动脉并发症，如扩张、动脉瘤和狭窄。虽然静脉注射免疫球蛋白（IVIG）能有效减轻免疫炎症，但10%-15%的患者对初始治疗无反应，甚至一些患者在连续两次治疗后仍表现出耐药性。预测哪些患者对这两种IVIG治疗无反应对于指导治疗策略和改善治疗结果至关重要。本研究旨在基于临床和实验室数据，使用先进的机器学习模型预测对连续两次IVIG治疗的耐药性。使用了韩国川崎病协会第9次全国川崎病患者调查的数据，该调查涵盖15378名患者（平均年龄33.0±24.8个月；性别比1.4:1）。临床和实验室检查结果包括白细胞计数、绝对中性粒细胞计数（ANC）、血小板计数、红细胞沉降率、血清蛋白、天冬氨酸转氨酶、丙氨酸转氨酶、总胆红素、N末端脑钠肽前体以及脓尿的存在情况。应用包括逻辑回归（LR）、多层感知器（MLP）、随机森林（RF）、CATBoost、可解释增强机器（EBM）和梯度增强机器（GBM）在内的机器学习模型来预测治疗耐药性。这些机器学习模型的受试者操作特征曲线下面积（AUROC）值在0.664至0.791之间，其中GBM模型的AUROC最高，为0.791。特征重要性分析表明，ANC、血清蛋白、血小板计数和C反应蛋白（CRP）水平是治疗耐药性的最显著预测因素。这些预测因素的截断值分别为：ANC为7860/mm³，血清蛋白为7.0 g/dL，血小板计数为519000/mm³，CRP为10.4 mg/dL。在这些患者中，12.2%对首次IVIG输注难治，2.8%对第二次IVIG治疗无反应。此外，这些患者中有13.1%在急性期确诊为冠状动脉扩张（CAD），4.7%在急性期后发生CAD。机器学习模型能有效预测对连续IVIG治疗的耐药性，有助于早期识别高危患者。关键预测因素包括ANC、血清蛋白、血小板计数和CRP水平。这些发现可指导个性化治疗策略并改善川崎病的治疗结果。
