基于自动机器学习的耳鸣中医辨证分型及关键因素研究

Traditional Chinese medicine syndrome differentiation and key factors of tinnitus based on automatic machine learning

目的应用自动机器学习技术构建耳鸣中医辨证分型模型并探索影响耳鸣辨证结果的关键因素。方法回顾性分析2021年1月至2022年1月上海市7家医疗单位收集的594例主观性耳鸣患者的临床特征,利用Auto-sklearn自动机器学习方法进行常见的15种机器学习算法对比,选择分类效果最优的模型分析影响耳鸣的关键因素。结果分类结果最优的算法是随机森林,它的准确率、精确度、敏感度、特异度、F1分数、曲线下面积(AUC)值、kappa系数分别为87.37%、88.34%、89.06%、96.63%、88.38%、97.50%、83.37%,并得出影响耳鸣类型肾精亏损、肝火上扰、痰火郁结、脾胃亏虚、风热侵袭分类的关键因素分别为滑脉、弦脉、滑脉、舌淡、浮脉。结论随机森林算法对于结构化的临床数据特征能提供很好的分类预测功能,提示机器学习技术对辅助中医耳鸣的诊断具有临床应用价值。

Objective To construct a traditional Chinese medicine syndrome differentiation model for tinnitus using automatic machine learning technology,and to explore the key factors that affect the results of tinnitus syndrome differentiation.Methods The clinical characteristics of 594 patients with subjective tinnitus in seven medical units in Shanghai from January 2021 to January 2022 were retrospectively analyzed.The Auto-sklearn automatic machine learning method was used to compare 15 algorithms,and the model with the best classification effect was selected to analyze the key factors affecting tinnitus.Results The results showed that the optimal algorithm for classification results was the random forest,its accuracy,precision,sensitivity,specificity,F1-score,AUC and kappa coefficient were 87.37%,88.34%,89.06%,96.63%,88.38%,97.50%,and 83.37%,respectively.It is concluded that the key factors affecting the classification of the pattern of kidney yin deficiency and fire effulgence,the pattern of liver fire disturbing upward,the pattern of stagnation and binding of phlegm and fire,the pattern of spleen and stomach deficiency,the pattern of wind and heat attacking the external are smooth pulse,string pulse,smooth pulse,weak tongue,and floating pulse respectively.Conclusions Random forest can provide a good classification prediction function for structured clinical data,suggesting that machine learning technology has clinical application value in assisting the diagnosis of subjective tinnitus.