慢性阻塞性肺疾病的中医智能诊疗研究

Study on TCM Intelligent Diagnosis and Treatment for Chronic Obstructive Pulmonary Disease

目的探讨深度学习在慢性阻塞性肺疾病证型预测和药物推荐中的应用。方法从真实诊疗数据中提取症状、证型、药物信息并做预处理,使用Fisher特征选择算法筛选与证型相关性较强的症状作为4层深度前馈网络的输入进行中医证型预测。将药物推荐分为典型药物推荐和补充药物推荐,利用频数分析和粒子群优化(PSO)算法-反向传播(BP)算法构建各证型的典型药物推荐模型,通过挖掘关联规则完成补充药物推荐。最后通过相应指标分别对证型预测、药物推荐结果进行评价。结果对2232条COPD患者数据的9种中医证型分类准确率达到82.39%。对于外寒内饮证,基于233种药物的典型药物推荐结果的均方误差(MSE)为0.0091,平均绝对误差(MAE)为0.0879。设置最小支持度0.2、最小置信度0.9,挖掘到关联规则261条,用于补充药物推荐。结论实验和实际使用结果表明,本研究提出的基于深度前馈网络的证型分类算法、基于频数分析和PSO-BP网络的药物推荐算法较好地完成COPD患者的证型预测及药物推荐,具有较好的智能诊疗效果。

Objective To explore the application of deep learning in syndrome prediction and medicine recommendation of chronic obstructive pulmonary disease(COPD).Methods The information of symptoms,syndrome types and medicine was extracted from real diagnosis and treatment data and was under preprocessing.Fisher feature selection algorithm was used to screen the strong correlation symptoms with syndrome types as the input of four-layer deep feedforward network to predict TCM syndrome types.The medicine recommendation was divided into typical medicine recommendation and supplementary medicine recommendation.The typical medicine recommendation models of each syndrome type were constructed using frequency analysis and PSO(particle swarm optimization)-BP(backpropagation algorithm)network,and the supplementary medicine recommendation was completed by mining association rules.Finally,the results of syndrome type prediction and medicine recommendation were evaluated by corresponding indicators.Results The classification accuracy of nine TCM syndrome types on 2232 COPD patient data was 82.39%.For syndrome of external cold and internal retained morbid fluid,the mean square error(MSE)of the recommended results of typical medicines based on 233 medicines was 0.0091,and the mean absolute error(MAE)was 0.0879.Totally 261 association rules for supplementary medicine recommendation have been mined when the minimum support and the minimum confidence were respectively set to 0.2 and 0.9.Conclusion The experimental and practical results show that the syndrome type classification algorithm based on deep feedforward network and the medicine recommendation algorithm based on frequency analysis and PSO-BP network proposed in this study can better complete the syndrome type prediction and medicine recommendation of COPD patients,and have better intelligent diagnosis and treatment effect.