预测抗结核药物性肝损伤的卷积神经网络与logistic回归模型的建立与比较

Development and comparison of convolutional neural network and logistic regression models for predicting anti-tuberculosis drug-induced liver injury

目的基于卷积神经网络(CNN)和多重logistic回归建立抗结核药物性肝损伤(ATB-DILI)的预测模型,评价和比较2个模型的性能。方法收集2019年1月1日至2022年10月31日镇江市第三人民医院、句容市人民医院和丹阳市第三人民医院的结核病住院患者临床及实验室检查数据。根据是否发生ATB-DILI将患者分为发生和未发生ATB-DILI组,比较2组临床特征差异。采用随机数字表法将患者按7︰3的比例分为训练集和测试集。基于训练集数据,分别采用多重logistic回归和CNN建立ATB-DILI预测模型;基于训练集和测试集数据,对2个模型预测ATB-DILI的准确度进行验证。绘制受试者工作特征(ROC)曲线,比较2种模型的灵敏度、特异度、约登指数和曲线下面积(AUC)。结果共有3012例患者纳入研究,其中294例(9.76%)被诊断为ATB-DILI;训练集2108例,测试集904例。多重logistic回归分析的结果显示,年龄、肝病史、低白蛋白血症和未预防性使用保肝药是ATB-DILI发生的独立风险因素,并以独立风险因素构建多重logistic回归模型方程。采用CNN对训练集患者数据的深度学习和分析结果显示,对ATB-DILI发生影响最大的前5位风险因素分别为肝病史、年龄、未预防使用保肝药、低白蛋白血症及饮酒,以前5位风险因素构建CNN模型。采用多重logistic回归模型预测的训练集与测试集患者ATB-DILI发生的总准确率分别为87.62%和88.27%;采用CNN模型预测训练集与测试集患者ATB-DILI发生的总准确率分别为92.36%和91.70%。CNN模型的灵敏度、特异度和AUC均高于多重logistic回归模型,差异均有统计学意义(均P<0.05)。结论多重logistic回归模型和CNN模型对ATB-DILI发生的预测性能均较好。但相较多重logistic回归模型,CNN模型对结核病住院患者ATB-DILI发生的预测性能更佳。

Objective To develop 2 prediction models for anti-tuberculosis drug-induced liver injury(ATB-DILI)based on convolutional neural network(CNN)and multiple logistic regression,and to evaluate and compare the performance of the 2 models.Methods The clinical and laboratory test data of inpatients in the Third People′s Hospital of Zhenjiang,Jurong People's Hospital,and the Third People′s Hospital of Danyang from January 1,2019 to October 31,2022 were collected.According to whether ATB-DILI occurred,patients were divided into with and without ATB-DILI groups,and the clinical characteristics of the 2 groups were compared.The patients were randomly divided into training set and test set according to a ratio of 7∶3 by random number table method.Based on data in the training set,multiple logistic regression and CNN were used to develop ATB-DILI prediction models;based on data in the training and test sets,the accuracy of the 2 models in predicting ATB-DILI was verified.The receiver operating characteristic(ROC)curve was drawn,and the sensitivity,specificity,Youden index and area under the curve(AUC)of the 2 models were compared.Results A total of 3-012 patients were included in the study,of which 294(9.76%)were diagnosed with ATB-DILI;2-108 patients were in the training set and 904 in the test set.The results of multiple logistic regression analysis showed that age,history of liver diseases,hypoalbuminemia,and no preventive use of liver protection drugs were independent risk factors for the occurrence of ATB-DILI.Based on these risk factors,multiple logistic regression model equations were constructed.The results of deep learning and analyzing the patient data of the training set by CNN showed that the top 5 risk factors that had the greatest impact on the occurrence of ATB-DILI were history of liver disease,age,no preventive use of liver protection drugs,hypoalbuminemia,and alcohol consumption.The CNN model was constructed according to the top 5 risk factors.The total accuracy in predicting the occurrence of ATB-DILI in the training and test sets using the multiple logistic regression model was 87.62%and 88.27%,respectively,and the total accuracy of using CNN model was 92.36%and 91.70%,respectively.The sensitivity,specificity,and AUC of the CNN model were all higher than those of the multiple logistic regression model,and the differences were statistically significant(all P<0.05).Conclusion Both the multiple logistic regression model and CNN model have good predictive performance for the occurrence of ATB-DILI,and the prediction performance of CNN model is better,comparatively.