基于神经网络分析方法探索类风湿关节炎证病信息对疗效的预测作用

目的:利用神经网络模型分析方法探索类风湿性关节炎(rheumatoidarthritis,RA)证候疾病信息对疗效的预测作用。方法:397例确诊为活动期RA的病例来自9个临床中心,随机分成中药治疗组203例和西药治疗组194例。西药治疗方案包括非甾体抗炎药和慢作用药,中药治疗包括基础治疗和辨证用药。治疗前后收集患者中医四诊信息和西医疾病诊查指标,治疗效果用美国风湿病学会20%改善标准(theAmericanCollegeofRheu-matology20,ACR20)判断,抽取患者初诊时的信息进行分析,分析方法在SAS8.2上实施。通过单因素探索性分析,计算疗效与变量的比数比,以P<0.2作为入选模型的标准;采用神经网络模型进行数据分析,以疗效为分层变量,随机将数据集分为训练集(占75%)和验证集(占25%),对分析方法进行验证。结果:数据分析模型中,中药治疗组共纳入18个变量,西药治疗组纳入24个变量。中药治疗组中,晨僵、关节肿胀数、免疫球蛋白M、关节压痛数、关节压痛、类风湿因子、C反应蛋白和关节疼痛等与疗效正相关,病程和夜尿多与疗效呈负相关。西药治疗组中,血沉、腰膝酸软、苔白、关节疼痛、屈伸不利和关节肿胀等与疗效呈正相关,苔黄、舌红、白细胞检测与疗效呈负相关。在随机选取的验证集患者中,神经网络模型的预测作用表现为:RA中药治疗方案使用中,可预测20%患者有效率达到90%;RA西药治疗方案使用中,可预测20%患者有效率达到100%。结论:根据证候疾病信息与中西医疗法疗效的临床数据所建立的神经网络模型,能够显示证病信息对疗效的预测作用。

Predictive role of diagnostic information in treatment efficacy of rheumatoid arthritis based on neural network model analysis

Objective: To analyze the indications of the therapies for rheumatoid arthritis (RA) with neural network model analysis. Methods: Three hundred and ninety-seven patients were included in the clinical trial from 9 clinical centers. They were randomly divided into Western medicine (WM) treated group, 194 cases; and traditional Chinese herbal medicine (CM) treated group, 203 cases. A complete physical examination and 18 common clinical manifestations were prepared before the randomization and after the treatment. The WM therapy included voltaren extended action tablet, methotrexate and sulfasalazine. The CM therapy included Glucosidorum Tripterygii Totorum Tablet and syndrome differentiation treatment. The American College of Rheumatology 20 (ACR20) was taken as efficacy evaluation. All data were analyzed on SAS 8.2 statistical package. The relationships between each variable and efficacy were analyzed, and the variables with P<0.2 were included for the data mining analysis with neural network model. All data were classified into training set (75%) and verification set (25%) for further verification on the data-mining model. Results: Eighteen variables in CM and 24 variables in WM were included in the data-mining model. In CM, morning stiffness, swollen joint number, peripheral immunoglobulin M (IgM) level, tenderness joint number, tenderness, rheumatoid factor (RF), C-reactive protein (CRP) and joint pain were positively related to the efficacy, and disease duration and more urination at night negatively related to the efficacy. In WM, erythrocyte sedimentation rate (ESR), weak waist, white fur in tongue, joint pain, joint stiffness and swollen joint were positively related to the efficacy, and yellow fur in tongue, red tongue, white blood negatively related to the efficacy. In the analysis with the neural network model in the patients of verification set, the predictive response rates of 20% patients would be 100% and 90% in the treatment with CM and WM, respectively. Conclusion: Neural network model analysis, based on the full clinical trial data with collection of both traditional Chinese medicine and modern medicine diagnostic information, shows a good predictive role for the information in the efficacy in rheumatoid arthritis.