Construction and optimization of traditional Chinese medicine constitution prediction models based on deep learning

Abstract

Objective

To cater to the demands for personalized health services from a deep learning perspective by investigating the characteristics of traditional Chinese medicine (TCM) constitution data and constructing models to explore new prediction methods.

Methods

Data from students at Chengdu University of Traditional Chinese Medicine were collected and organized according to the 24 solar terms from January 21, 2020, to April 6, 2022. The data were used to identify nine TCM constitutions, including balanced constitution, Qi deficiency constitution, Yang deficiency constitution, Yin deficiency constitution, phlegm dampness constitution, damp heat constitution, stagnant blood constitution, Qi stagnation constitution, and specific-inherited predisposition constitution. Deep learning algorithms were employed to construct multi-layer perceptron (MLP), long short-term memory (LSTM), and deep belief network (DBN) models for the prediction of TCM constitutions based on the nine constitution types. To optimize these TCM constitution prediction models, this study introduced the attention mechanism (AM), grey wolf optimizer (GWO), and particle swarm optimization (PSO). The models’ performance was evaluated before and after optimization using the F1-score, accuracy, precision, and recall.

Results

The research analyzed a total of 31 655 pieces of data. (i) Before optimization, the MLP model achieved more than 90% prediction accuracy for all constitution types except the balanced and Qi deficiency constitutions. The LSTM model's prediction accuracies exceeded 60%, indicating that their potential in TCM constitutional prediction may not have been fully realized due to the absence of pronounced temporal features in the data. Regarding the DBN model, the binary classification analysis showed that, apart from slightly underperforming in predicting the Qi deficiency constitution and damp heat constitution, with accuracies of 65% and 60%, respectively. The DBN model demonstrated considerable discriminative power for other constitution types, achieving prediction accuracy rates and area under the receiver operating characteristic (ROC) curve (AUC) values exceeding 70% and 0.78, respectively. This indicates that while the model possesses a certain level of constitutional differentiation ability, it encounters limitations in processing specific constitutional features, leaving room for further improvement in its performance. For multi-class classification problem, the DBN model’s prediction accuracy rate fell short of 50%. (ii) After optimization, the LSTM model, enhanced with the AM, typically achieved a prediction accuracy rate above 75%, with lower performance for the Qi deficiency constitution, stagnant blood constitution, and Qi stagnation constitution. The GWO-optimized DBN model for multi-class classification showed an increased prediction accuracy rate of 56%, while the PSO-optimized model had a decreased accuracy rate to 37%. The GWO-PSO-DBN model, optimized with both algorithms, demonstrated an improved prediction accuracy rate of 54%.

Conclusion

This study constructed MLP, LSTM, and DBN models for predicting TCM constitution and improved them based on different optimisation algorithms. The results showed that the MLP model performs well, the LSTM and DBN models were effective in prediction but with certain limitations. This study also provided a new technology reference for the establishment and optimisation strategies of TCM constitution prediction models,and a novel idea for the treatment of non-disease.