Takeuchi Masato, Inuzuka Ryo, Hayashi Taiyu, Shindo Takahiro, Hirata Yoichiro, Shimizu Nobutaka, Inatomi Jun, Yokoyama Yoshiki, Namai Yoshiyuki, Oda Yoichiro, Takamizawa Masaru, Kagawa Jiro, Harita Yutaka, Oka Akira
From the *Department of Pediatrics, Kikkoman General Hospital, Chiba, Japan; †Department of Pediatrics, The University of Tokyo, Tokyo, Japan; ‡Department of Pediatrics, Yaizu City Hospital, Shizuoka, Japan; §Department of Pediatrics, Ome Municipal Hospital, Tokyo, Japan; ¶Department of Pediatrics, Ohta-Nishinouchi Hospital, Fukushima, Japan; ‖Department of Pediatrics, Chigasaki Municipal Hospital, Kanagawa, Japan; **Department of Pediatrics, Saitama Citizens Medical Center, Saitama, Japan; and ††Department of Pediatrics, Fujieda Municipal General Hospital, Shizuoka, Japan.
Pediatr Infect Dis J. 2017 Sep;36(9):821-826. doi: 10.1097/INF.0000000000001621.
Resistance to intravenous immunoglobulin (IVIG) therapy is a risk factor for coronary lesions in patients with Kawasaki disease (KD). Risk-adjusted initial therapy may improve coronary outcome in KD, but identification of high risk patients remains a challenge. This study aimed to develop a new risk assessment tool for IVIG resistance using advanced statistical techniques.
Data were retrospectively collected from KD patients receiving IVIG therapy, including demographic characteristics, signs and symptoms of KD and laboratory results. A random forest (RF) classifier, a tree-based machine learning technique, was applied to these data. The correlation between each variable and risk of IVIG resistance was estimated.
Data were obtained from 767 patients with KD, including 170 (22.1%) who were refractory to initial IVIG therapy. The predictive tool based on the RF algorithm had an area under the receiver operating characteristic curve of 0.916, a sensitivity of 79.7% and a specificity of 87.3%. Its misclassification rate in the general patient population was estimated to be 15.5%. RF also identified markers related to IVIG resistance such as abnormal liver markers and percentage neutrophils, displaying relationships between these markers and predicted risk.
The RF classifier reliably identified KD patients at high risk for IVIG resistance, presenting clinical markers relevant to treatment failure. Evaluation in other patient populations is required to determine whether this risk assessment tool relying on RF has clinical value.
对静脉注射免疫球蛋白（IVIG）治疗耐药是川崎病（KD）患者发生冠状动脉病变的一个危险因素。风险调整后的初始治疗可能会改善KD患者的冠状动脉结局，但识别高危患者仍然是一项挑战。本研究旨在使用先进的统计技术开发一种用于评估IVIG耐药的新风险评估工具。
回顾性收集接受IVIG治疗的KD患者的数据，包括人口统计学特征、KD的体征和症状以及实验室检查结果。将一种基于树的机器学习技术——随机森林（RF）分类器应用于这些数据。估计每个变量与IVIG耐药风险之间的相关性。
从767例KD患者中获取数据，其中170例（22.1%）对初始IVIG治疗无效。基于RF算法的预测工具的受试者工作特征曲线下面积为0.916，灵敏度为79.7%，特异度为87.3%。估计其在普通患者群体中的误分类率为15.5%。RF还识别出了与IVIG耐药相关的标志物，如肝脏标志物异常和中性粒细胞百分比，并显示了这些标志物与预测风险之间的关系。
RF分类器能够可靠地识别出有IVIG耐药高风险的KD患者，并呈现出与治疗失败相关的临床标志物。需要在其他患者群体中进行评估，以确定这种基于RF的风险评估工具是否具有临床价值。
