Application of machine learning in Chinese medicine differentiation of dampness-heat pattern in patients with type 2 diabetes mellitus

Abstract

Background: China has become the country with the largest number of people with type 2 diabetes mellitus (T2DM), and Chinese medicine (CM) has unique advantages in preventing and treating T2DM, while accurate pattern differentiation is the guarantee for proper treatment.

Objective: The establishment of the CM pattern differentiation model of T2DM is helpful to the pattern diagnosis of the disease. At present, there are few studies on dampness-heat pattern differentiation models of T2DM. Therefore, we establish a machine learning model, hoping to provide an efficient tool for the pattern diagnosis of CM for T2DM in the future.

Methods: A total of 1021 effective samples of T2DM patients from ten CM hospitals or clinics were collected by a questionnaire including patients' demographic and dampness-heat-related symptoms and signs. All information and the diagnosis of the dampness-heat pattern of patients were completed by experienced CM physicians at each visit. We applied six machine learning algorithms (Artificial Neural Network [ANN], K-Nearest Neighbor [KNN], Naïve Bayes [NB], Support Vector Machine [SVM], Extreme Gradient Boosting [XGBoost] and Random Forest [RF]) and compared their performance. And then we also utilized Shapley additive explanation (SHAP) method to explain the best performance model.

Results: The XGBoost model had the highest AUC (0.951, 95% CI 0.925-0.978) among the six models, with the best sensitivity, accuracy, F1 score, negative predictive value, and excellent specificity, precision, and positive predictive value. The SHAP method based on XGBoost showed that slimy yellow tongue fur was the most important sign in dampness-heat pattern diagnosis. The slippery pulse or rapid-slippery pulse, sticky stool with ungratifying defecation also performed an important role in this diagnostic model. Furthermore, the red tongue acted as an important tongue sign for the dampness-heat pattern.

Conclusion: This study constructed a dampness-heat pattern differentiation model of T2DM based on machine learning. The XGBoost model is a tool with the potential to help CM practitioners make quick diagnosis decisions and contribute to the standardization and international application of CM patterns.

Keywords: Dampness-heat pattern; Diagnostic model; Machine learning; Pattern differentiation.

Fig. 1

Flow diagram of patient selection and model building.

Fig. 2a

Evaluation of the six machine learning algorithms based on the AUC of the ROC curve. AUC, area under the curve; ROC, receiver operating characteristic; 2b P-R curve and AP of the six models. P-R curve: precision-recall curve; AP: average precision.

Fig. 3

Shapley additive explanation (SHAP) values to show the interpretability of the effects of variables as the input factors for the diagnosis of the dampness-heat pattern of T2DM. (a) SHAP values for all 1021 patients in the train set. (b, c) SHAP values of two typical patients from the positive group (b) and the negative group (c) are illustrated with their most important variables.

Fig. 4a

Importance matrix plot of the XGBoost model, depicting the importance of each symptom and signs for diagnosing dampness-heat pattern in T2DM.4b SHAP summary plot of the 14 symptoms and signs of the XGBoost model. There is one dot per patient per symptom or sign colored according to an attribution value, where red represents a higher value and blue represents a lower value.