基于机器学习模型预测经皮肾镜取石术后全身炎症反应综合征

Machine Learning Models to Predict Systemic Inflammatory Response Syndrome after Percutaneous Nephrolithotomy

摘要: 目的：本研究旨在开发和评估机器学习模型在预测经皮肾镜取石术(Percutaneous nephrolithotomy，PCNL)术后发生全身炎症反应综合征(Systemic inflammatory response syndrome，SIRS)的性能。方法：回顾性分析2021年10月至2022年12月接受PCNL治疗的218例患者的临床资料。在我们的研究中，按照8:2划分数据集为训练集和测试集。基于Light Gradient Boosting Machine (LightGBM) 机器学习算法在训练集构建预测模型。LightGBM机器学习模型的预测性能由使用测试集的受试者工作特征曲线下面积(Area under the receiver operating characteristic curve，AUC)、准确性、灵敏度和特异性决定。我们使用权重系数来解释每个变量对预测模型性能的贡献。结果：LightGBM模型在测试集中的准确率为0.837，AUC为0.918 (95%CI 0.827-1.000)，灵敏度为0.875，特异性为0.829。对LightGBM模型的进一步分析显示，全身免疫炎症(systemic immune-inflammation，SII)指标对结局预测的贡献最大，其次是术前尿培养、前白蛋白、结石负荷、中性粒细胞与淋巴细胞比值(neutrophil to lymphocyte ratio，NLR)、肾盂积水、纤维蛋白原、淋巴细胞与单核细胞比值(Lymphocyte to monocyte ratio，LMR)。结论：基于LightGBM模型学习患者临床数据，能够提前准确预测PCNL术后SIRS发生的可能性，并用于指导外科医生的临床决策。

Abstract: Objective: The objective of this study was to develop and evaluate the performance of machine learning model for predicting the possibility of systemic inflammatory response syndrome (SIRS) following percutaneous nephrolithotomy (PCNL). Methods: We retrospectively reviewed the clinical data of 218 patients who received PCNL between October 2021 and December 2022. In our study, the dataset is divided into the training set and the testing set according to 8:2. The prediction mod-el based on the Light Gradient Boosting Machine (LightGBM) algorithms was created using the training set. The predictive performance of the LightGBM machine learning model was determined by the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and spec-ificity using the testing set. We used coefficients to interpret the contribution of each variable to the predictive performance. Results: The LightGBM model delivered a good performance with an accu-racy of 0.837, AUC of 0.918 (95% CI 0.827~1.000), sensitivity of 0.875, specificity of 0.829 in the testing set. Further analysis using the LightGBM model showed that systemic immune inflammation (SII) contributed the most to the prediction of the outcome, followed by preoperative urine culture, prealbumin, stone burden, neutrophil to lymphocyte ratio (NLR), hydronephrosis, fibrinogen, and Lymphocyte to monocyte ratio (LMR). Conclusion: The LightGBM models can accurately predict the possibility of SIRS after PCNL in advance by learning patient clinical data, and should be used to guide surgeons in clinical decision-making.