Data Mining of Tongue and Pulse based on Tongue Diagnosis and Pulse Diagnosis Instruments：Evaluation of Machine-Learning Assisting Classification of Fatigue Subjects (Preprint)

BACKGROUND Fatigue is one of the most common subjective symptom,due to lack of objective diagnostic criteria ,it is often neglected,especially in the early stage of disease.It is very important to establish an objective evaluation method for diagnosis of disease fatigue and non-disease fatigue based on tongue and pulse data. Our research team has been engaged in the informatization and intelligentization research of TCM diagnosis technology for more than 30 years, we have been engaged in health evaluation and analysis of large clinical sample data for many years, and accumulated a certain amount of typical data,in this area of research, foundation of our work has outstanding advantages. OBJECTIVE The purposes of this study is to explore a new method for clinical diagnosis of fatigue, by analyzing the characteristic of objective tongue and pulse data of disease fatigue subjects and sub-health fatigue subjects, and establishing the diagnosis and classification models of sub-health fatigue and disease fatigue based on the tongue and pulse data. METHODS In this study,736 subjects were selected from the Physical Examination Center, and were divided into healthy control, sub-health fatigue group and disease fatigue group according to physical examination indexes and health status assessment criteria. TFDA-1 tongue diagnosis instrument and PDA-1 pulse diagnosis instrument independently developed were used to collect the tongue image and sphygmogram of the subjects. Multivariate statistical analysis methods were usded to analyze the tongue and pulse data distribution pattern of the fatigue subjects, and machine learning methods were used to establish the diagnostic classification models of fatigue. RESULTS The tongue feature TB-a,TB-b,TB-H,TB-S,TB-I,TB-Cb,TC-L,TC-H,TC-I,TC-Y,TC-Cr,TC-Cb,perAll and perPart were statistically significant different among three groups (P < 0.05, P < 0.01). The sphygmogram feature t1,t4,h5,W1,W2,W1/t and W2/t showed statistically significant differences among three groups (P < 0.05, P < 0.01). There was a canonical correlation between the tongue and pulse data in the disease fatigue group, and the correlation coefficient was 0.649 (P<0.05). Logistic Regression models showed good classification efficiency for data sets of " tongue ", " basic information "and "tongue + pulse + basic information". The area under ROC curve was 0.7352, 0.9089 and 0.9460, respectively. And the Neural Network models showed better classification efficiency for data sets of " pulse " and "tongue + pulse", the area under ROC curve was 0.7957 and 0.8293, respectively. Besides Logistic Regression and Neural Network, Random Forest and SVM also showed good classification results in the diagnosis models. And the classification efficiency of models based on different data sets had the following trends in the three groups: "tongue"models <"pulse" models <" tongue+pulse"models <" basic information"models <" tongue+pulse+basic information "models. CONCLUSIONS Sub-health fatigue and disease fatigue subjects had different tongue and pulse characteristic. Diagnosis and classification models can be established using basic information, tongue and pulse data to explore a new method for evaluating fatigue state.