目的：应用支持向量机（support vector machine，SVM）构建ICU中急性肾功能损伤(acute kidney injury, AKI)患者住院死亡风险预测模型，并比较其与ICU中常用的简化急性生理评分（the simplified acute physiology score Ⅱ，SAPS-Ⅱ）的预测性能。方法：使用重症监护医学信息市场（medical information mart for intensive care Ⅲ，MIMIC-Ⅲ）数据库作为数据来源。根据2012年国际改善全球肾脏病预后组织（Kidney Disease: Improving Global Outcomes, KDIGO）发表的《急性肾损伤临床实践指南》选取MIMICⅢ数据库中的AKI患者，使用SAPS-Ⅱ中所用到的全部变量构建SVM模型,同时，使用MIMIC-Ⅲ数据库定制本地化的SAPSⅡ模型，并比较其与SVM模型的性能优劣。模型性能的评价方法使用五折交叉验证，评价指标使用受试者工作特征曲线下面积(area under the receiver operation characteristic curve，AUROC)、均方根误差（root mean squared error，RMSE）、灵敏度、特异度和准确率。此外，使用Bland-Altman图评估两模型预测结果的一致性。结果：共纳入19 044例AKI患者，死亡率为13.58%。五折交叉验证的结果显示，SVM模型和定制版SAPS-Ⅱ模型的平均AUROC分别为0.86和0.81, 差异有统计学意义（t=13.0,P<0.001）， SVM模型和定制版SAPS-Ⅱ模型的平均RMSE分别为0.29和0.31, 差异有统计学意义（t=-9.6,P<0.001）。在灵敏度和约登指数方面，SVM模型也均优于定制版的SAPS-Ⅱ模型，差异均具有统计学意义（P分别为0.002和<0.001）。BlandAltman图显示当患者死亡风险极高或者极低时，两模型预测结果的一致性较好。当患者死亡风险的不确定性较大时，两模型预测结果的一致性较差。结论：相比于传统的SAPS-Ⅱ模型，SVM模型的预测性能更优，且当患者的死亡风险不确定时，这种优势尤其明显; SVM模型更有利于AKI患者的死亡风险识别与早期干预，能有效地帮助ICU临床医生提高医疗质量，有很强的临床应用价值。
Objective： To construct an in-hospital mortality prediction model for patients with acute kidney injury (AKI) in intensive care unit (ICU) by using support vector machine (SVM), and compare it with the simplified acute physiology score Ⅱ (SAPS-Ⅱ) which is commonly used in the ICU. Methods： We used Medical Information Mart for Intensive Care Ⅲ (MIMIC-Ⅲ) database as data source. The AKI patients in the MIMIC-Ⅲ database were selected according to the 2012 Kidney Disease: Improving Global Outcomes (KDIGO) definition of AKI. We employed the same predictor variable set as used in SAPS-Ⅱ to construct an SVM model. Meanwhile, we also developed a customized SAPS-Ⅱ mo-del using MIMIC-Ⅲ database, and compared performances between the SVM model and the customized SAPS-Ⅱ model. The performance of each model was evaluated via area under the receiver operation characteristic curve (AUROC), root mean squared error (RMSE), sensitivity, specificity, Youden’s index and accuracy based on 5-fold cross-validation. The agreement of the results between the SVM mo-del and the customized SAPS-Ⅱ model was illustrated using Bland-Altman plots. Results: A total number of 19 044 patients with AKI were included. The observed in-hospital mortality of the AKI patients was 13.58% in MIMIC-Ⅲ. The results based on the 5-fold cross validation showed that the average AUROC of the SVM model and the customized SAPS-Ⅱ model was 0.86 and 0.81, respectively (The difference between the two models was statistically significant with t=13.0, P<0.001). The average RMSE of the SVM model and the customized SAPS-Ⅱ model was 0.29 and 0.31, respectively (The difference was statistically significant with t=-9.6, P<0.001). The SVM model also outperformed the customized SAPS-Ⅱ model in terms of sensitivity and Youden’s index with significant statistical differences (P = 0.002 and <0.001, respectively).The Bland-Altman plot showed that the SVM model and the custo-mized SAPS-Ⅱ model had similar mortality prediction results when the mortality of a patient was certain, but the consistency between the mortality prediction results of the two models was poor when the mortality of a patient was with high uncertainty. Conclusion: Compared with the SAPS-Ⅱ model, the SVM model has a better performance, especially when the mortality of a patient is with high uncertainty. The SVM model is more suitable for predicting the mortality of patients with AKI in ICU and early intervention in patients with AKI in ICU. The SVM model can effectively help ICU clinicians improve the quality of me-dical treatment, which has high clinical value.
