[Comparison of machine learning method and logistic regression model in prediction of acute kidney injury in severely burned patients].

Objective: To build risk prediction models for acute kidney injury (AKI) in severely burned patients, and to compare the prediction performance of machine learning method and logistic regression model. Methods: The clinical data of 157 severely burned patients in August 2nd Kunshan factory aluminum dust explosion accident conforming to the inclusion criteria were collected. Patients suffering AKI within 90 days after admission were enrolled in group AKI, while the others were enrolled in non-AKI group. Single factor analysis was used to choose independent factors associated with AKI, including sex, age, admission time, features of basic injuries, initial score on admission, treatment condition, and mortality on post injury days 30, 60, and 90. Data were processed with Mann-Whitney U test, chi-square test, and Fisher's exact test. Variables with P<0.1 in single factor analysis and those with possible clinical significance were brought into the establishment of prediction model. Logistic regression and XGBoost machine learning algorithm were used to build the prediction model of AKI. The area under receiver operating characteristic curve (AUC) was calculated, and the sensitivity and specificity for optimal threshold value were also calculated for each model. Nonparametric resampling test was used to compare the significance of difference of AUC of the two models. Results: (1) Eighty-nine (56.7%) patients developed AKI within 90 days from admission. Compared with 68 patients in non-AKI group, 89 patients in group AKI were older (Z=-2.203, P<0.05), with larger total burn area and full-thickness burn area (Z=-5.200, -6.297, P<0.01), worse acute physical and chronic health evaluation (APACHE) Ⅱ score, abbreviated burn severity index score, and sequential organ failure assessment (SOFA) score on admission (Z=-7.485, -4.739, -4.590, P<0.01), higher occurrence rate of sepsis (χ(2)=33.087, P<0.01), higher rates of accepting tracheotomy, mechanical ventilation, and continuous renal replacement therapy (χ(2)=12.373, 17.201, 43.763, P<0.01), larger first excision area (Z=-2.191, P<0.05), and higher mortality on post injury days 30, 60, and 90 (χ(2)=7.483, 37.259, 45.533, P<0.01). There were no statistically significant differences in sex, open decompression, admission time, 24-hour fluid volume after admission, 48-hour fluid volume after admission, the first 24-hour urine volume, the second 24 hour urine volume, the first excision time, and inhalation injury (χ(2)=0.529, 3.318, Z=-1.746, -0.016, -1.199, -1.824, -0.625, -1.747, P>0.05). The rates of deep vein catheterization of patients in the two groups were both 100%. (2) There were twenty possible prediction variables for preliminary establishment of model according to the difference results of single factor analysis and clinical significance of variables. (3) The logistic regression prediction model had three variables: APACHE Ⅱ score [odds ratio (OR)=1.36, 95% confidence interval (CI)=1.20-1.53, P<0.001], sepsis (OR=2.63, 95% CI=0.90-7.66, P>0.05), and the first 24-hour urine volume (OR=0.71, 95% CI=0.50-1.01, P>0.05). The AUC of the logistic regression prediction model was 0.875 (95% CI=0.821-0.930), with the specificity and sensitivity of optimal threshold value 84.4% and 77.7%, respectively. (4) XGBoost machine learning model had seven main predictive variables: APACHE Ⅱ score, full-thickness burn area, 24-hour fluid volume after admission, sepsis, the first 24-hour urine volume, SOFA score, and 48-hour fluid volume after admission. The AUC of machine learning model was 0.920 (95% CI=0.879-0.962), higher than that of logistic regression model (P<0.001), with the specificity and sensitivity of optimal threshold value 89.7% and 82.0%, respectively. Conclusions: Sepsis and fluid resuscitation are two important predictive variables that can be intervened for AKI in severely burned patients. Machine learning method has a better performance and can provide more accurate prediction for individuals than logistic regression prediction model, and therefore has good clinical application prospect.目的： 构建严重烧伤患者发生急性肾损伤(AKI)的风险预测模型，比较机器学习和logistic回归模型的预测效能。 方法： 收集在"八二"昆山工厂铝粉尘爆炸事故中严重烧伤的符合入选标准的157例患者的临床资料。将入院90 d内发生AKI的患者纳入AKI组，其余患者纳入非AKI组。使用单因素分析筛选可能和AKI发生相关的因素，包括患者性别、年龄、入院耗时、基础伤情、入院初始评分、治疗情况以及伤后30、60、90 d病死率等指标。对数据行Mann－Whitney U检验、χ(2)检验、Fisher确切概率法检验。将单因素分析中P<0.1以及可能有临床意义的变量纳入预测模型的构建，分别采用logistic回归分析和XGBoost机器学习算法构建AKI预测模型。计算模型的受试者工作特征曲线下面积(AUC)，以及最佳阈值下的敏感度、特异度。2个预测模型AUC差异的显著性检验采用非参数的重复采样方法。 结果： (1)患者中89例(56.7%)在入院90 d内发生了AKI。与非AKI组68例患者相比，AKI组89例患者的年龄更大(Z＝－2.203，P<0.05)，烧伤总面积和Ⅲ度烧伤面积更大(Z＝－5.200、－6.297，P<0.01)，入院时的急性生理与慢性健康评估Ⅱ(APACHEⅡ)评分、简明烧伤严重程度指数评分、序贯器官衰竭评估(SOFA)评分更差(Z＝－7.485、－4.739、－4.590，P<0.01)，发生脓毒症的百分比更高(χ(2)＝33.087，P<0.01)，接受气管切开、呼吸机辅助呼吸以及连续性肾脏替代治疗的百分比更高(χ(2)＝12.373、17.201、43.763，P<0.01)，首次切痂面积更大(Z＝－2.191，P<0.05)，伤后30、60、90 d病死率更高(χ(2)＝7.483、37.259、45.533，P<0.01)。2组患者的性别、切开减张、入院耗时、入院后24 h补液量、入院后48 h补液量、第1个24 h尿量、第2个24 h尿量、首次切痂时间、吸入性损伤比较，差异无统计学意义(χ(2)＝0.529、3.318，Z＝－1.746、－0.016、－1.199、－1.824、－0.625、－1.747，P>0.05)；深静脉置管率均为100%。(2)根据单因素分析的差异结果以及变量的临床意义，筛选出20个可能的预测自变量，供模型的初步构建。(3)logistic回归预测模型的自变量为APACHEⅡ评分(比值比为1.36，95%置信区间为1.20～1.53，P<0.001)、脓毒症(比值比为2.63，95%置信区间为0.90～7.66，P>0.05)及第1个24 h尿量(比值比为0.71，95%置信区间为0.50～1.01，P>0.05)。logistic回归分析构建预测模型的AUC为0.875(95%置信区间为0.821～0.930)，最佳阈值下的特异度为84.4%、敏感度为77.7%。(4)XGBoost机器学习算法构建的模型有7个主要预测变量：APACHEⅡ评分、Ⅲ度烧伤面积、入院后24 h补液量、脓毒症、第1个24 h尿量、SOFA评分、入院后48 h补液量。机器学习算法构建预测模型的AUC为0.920(95%置信区间为0.879～0.962)，比logistic回归分析构建模型更高(P<0.001)，最佳阈值下的特异度为89.7%、敏感度为82.0%。 结论： 脓毒症和液体复苏情况是严重烧伤患者发生AKI的可干预的重要预测变量。机器学习模型预测严重烧伤患者发生AKI的性能较logistic回归预测模型更佳，能为患者提供更为精准的个体化预测，具有良好的临床应用前景。.