基于机器学习决策树模型对急性百草枯中毒患者预后的预测价值

Predictive value of decision tree-based machine learning model for prognosis in acute paraquat poisoning

目的探讨基于机器学习的决策树模型对急性百草枯(PQ)中毒(APP)患者预后的预测价值。方法采用回顾性研究方法。收集2012年5月至2021年8月沧州市中心医院急诊医学部救治的APP患者的临床数据,包括性别、年龄、服毒至洗胃的时间、血液灌流比例、血PQ浓度、生化指标[白细胞计数(WBC)丙氨酸转氨酶(ALT)血肌酐(SCr)血淀粉酶及血钾]以及血气指标[动脉血乳酸(Lac)、剩余碱和动脉血氧分压(PaO_(2))]。根据中毒后90d预后将患者分为生存组(56例)和死亡组(74例),比较不同预后两组患者临床指标的差异。通过多因素Logistic回归分析影响APP患者预后的危险因素,将危险因素作为变量构建含血PQ浓度及不含血PQ浓度的两种决策树模型。绘制受试者工作特征曲线(ROC曲线)评估决策树模型对APP患者预后的预测价值,通过Hanley&McNeil法对两种决策树模型的ROC曲线下面积(AUC)进行比较。结果患者90d的生存率为43.1%(56/130)。与死亡组比较,生存组WBC[×10^(9)/L:8.9(7.0,11.6)比17.4(11.9,23.1)]、ALT[U/L:25.3(21.2,31.8)比29.3(23.2,40.3)]、SCr[μmol/L:64.0(53.0,74.0)比91.0(72.5,141.5)]、Lac[mmol/L:2.5(1.4,4.0)比7.1(3.7,11.0)]和血PQ浓度[ng/L:0.3(0.10.9)比2.9(1.9,8.1)]均较低,差异均有统计学意义(均P<0.05),剩余碱[mmol/L:-2.5(-4.2,-1.1)比-7.2(-10.9,-4.7)]和血钾[mmol/L:3.7(3.5,4.0)比3.2(2.83.7)]均较高,差异均有统计学意义(均P<0.05),患者更年轻[岁:33.5(26.0,47.8)比42.5(26.0,58.0),P<0.05]。单因素Logistic回归分析结显示,年龄、WBC、ALT、SCr、血钾、Lac、剩余碱和血PQ浓度是影响APP患者90d预后的独立危险因素[优势比(0R)和95%可信区间(95%CI)分别为1.03(1.01~1.05)1.30(1.18~1.44)1.04(1.01~1.07)、1.02(1.01~1.04)、7.59(3.25~17.70)、1.64(1.35~1.99)、1.51(1.29~1.76)、7.00(3.41~14.37),P值分别为0.018、<0.001、0.011、<0.001、<0.001、<0.001、<0.001、<0.001]。含血PQ浓度的多因素Logistic回归分析结果显示,WBC、血钾和血PQ浓度是影响患者90d生存的独立危险因素[OR和95%CI分别为1.17(1.03~1.33)、7.29(1.66~32.01)、5.49(2.48~12.13),P值分别为0.014、0.008、<0.001]。不含血PQ浓度的多因素Logistic回归分析显示,年龄、WBC、血钾和剩余碱是影响患者90d生存的独立危险因素[OR和95%CI分别为1.05(1.01~1.08)、1.20(1.07~1.34)3.12(1.01~9.66)1.41(1.16~1.72),P值分别为0.008、0.002、0.049、0.001]。基于血PQ浓度和血钾决策树模型的AUC为0.94,95%CI为0.89~0.98,敏感度为91.9%,特异度为89.3%,准确率为90.0%。基于WBC、剩余碱和年龄决策树模型的AUC为0.89,95%CI为0.84~0.95,敏感度为86.5%,特异度为91.1%,准确率为88.5%。Hanley&McNeil法比较显示,两种决策树模型的AUC差异无统计学意义(Z=1.34,P=0.180)。结论基于机器学习的决策树模型可为临床早期评估APP患者的预后提供定量、直观的预测工具。

Objective To investigate the predictive value of a decision tree-based machine learning model for prognosis in acute paraquat(PQ)poisoning(APP)patients.Methods A retrospective study was conducted.The clinical data of APP patients from Cangzhou Central Hospital between May 2012 and August 2021 were collected,including gender,age,time from ingestion to gastric lavage,proportion of hemoperfusion,serum PQ concentration,biochemical indicators[white blood cell count(WBC),alanine aminotransferase(ALT),serum creatinine(SCr),serum amylase,and serum potassium],and blood gas indicators[arterial blood lactic acid(Lac),base excess(BE),and arterial partial pressure of oxygen(PaO_(2))].Patients were divided into a survival group(n=56)and a death group(n=74)based on 90-day prognosis,and the clinical data between the two groups were compared.The multivariate Logistic regression analysis was conducted to analyze the risk factors of prognosis in APP patients,and two decision tree models(i.e.,with/without serum PQ concentration)were constructed based on the risk factors.The predictive value was evaluated by the receiver operator characteristic(ROC)curve,and the area under the ROC curve(AUC)of two decision tree models was compared by Hanley&McNeil method.Results The 90-day survival rate of the patients was 43.1%(56/130).Compared with death group,patients in the survival group had lower WBC[×10^(9)/LL:8.9(7.0,11.6)vs.17.4(11.9,23.1)],ALT[U/L:25.3(21.2,31.8)vs.29.3(23.2,40.3)],SCr[μmol/L:64.0(53.0,74.0)vs.91.0(72.5,141.5)],Lac[mmol/L:2.5(1.4,4.0)vs.7.1(3.7,11.0)],and serum PQ concentration[ng/L:0.3(0.1,0.9)vs.2.9(1.9,8.1)],the difference were statistically significant(all P<0.05),higher BE[mmol/L:-2.5(-4.2,-1.1)vs.-7.2(-10.9,-4.7)]and serum potassium[mmol/L:3.7(3.5,4.0)vs.3.2(2.8,3.7)],the difference were statistically significant(all P<0.05),and patients were younger[years:33.5(26.0,47.8)vs.42.5(26.0,58.0),P<0.05].Univariate Logistic regression analysis showed that age,WBC,ALT,SCr,serum potassium,Lac,BE and serum PQ concentration were independent risk factors of 90-day survival[odds ratio(0R)and 95%confidence interval(95%CI) were 1.03(1.01-1.05),1.30(1.18-1.44),1.04(1.01-1.07),1.02(1.01-1.04),7.59(3.25-17.70),1.64(1.35-1.99),1.51(1.29-1.76),7.00(3.41-14.37),P values were 0.018,<0.001,0.011,<0.001,<0.001,<0.001,<0.001,<0.001].Multivariate Logistic regression analysis with serum PQ concentration showed that WBC,serum potassium,and serum PQ concentration were independent risk factors for 90-day survival[0R and 95%CI were 1.17(1.03-1.33),7.29(1.66-32.01),5.49(2.48-12.13),P values were 0.014,0.008,<0.001].Multivariate Logistic regression analysis without serum PQ concentration showed that age,WBC,serum potassium and BE were independent risk factors for 90-day survival[0R and 95%CI were 1.05(1.01-1.08),1.20(1.07-1.34),3.12(1.01-9.66),1.41(1.16-1.72),P values were 0.008,0.002,0.049,0.001].The decision tree model based on serum PQ concentration and serum potassium showed an AUC of 0.94(95%CI was 0.89-0.98),along with 91.9%sensitivity,89.3%specificity,and 90.0%accuracy.The decision tree model based on WBC,BE,and age showed an AUC of 0.89(95%Cl was 0.84-0.95),with 86.5%sensitivity,91.1%specificity,and 88.5%accuracy.Pairwise comparison of the AUC using Hanley&McNeil method demonstrated that no statistical difference between the two decision tree models(Z=1.34,P=0.180).Conclusion The decision tree-based models can provide quantitative and intuitive prediction tools for the early detection of prognosis in APP patients in clinical practice.