Unsupervised clustering of concurrent constitutions in traditional Chinese medicine for personalized treatment

Abstract The complexity of treating concurrent constitutions in Traditional Chinese Medicine (TCM) stems from 248 possible combinations, posing significant challenges for personalized treatment. This study applies unsupervised machine learning to identify data-derived symptom subgroups within mixed TCM constitutions, providing a framework for characterizing heterogeneity. A large-scale questionnaire dataset based on the Constitution in Chinese Medicine Questionnaire (CCMQ) was analyzed. Both binary-encoded symptom data and original 5-point Likert-scale data were considered. We employed three dimensionality reduction techniques (UMAP, t-SNE, PCA) and three clustering algorithms (K-means, DBSCAN, GMM) to analyze symptom-based datasets encoded in both a 5-point Likert-scale and binary format. Cluster validity was assessed using the Silhouette Score, Calinski-Harabasz Score, and Davies-Bouldin Score. Across different analytical settings, the combination of binary encoding, UMAP, and Gaussian mixture modeling showed comparatively favorable and stable performance among the tested pipelines. Under this framework, a ten-cluster solution was identified, with clusters exhibiting distinguishable symptom patterns, including heat-related, cold-sensitive, and mixed physiological profiles. These patterns were generally consistent across sensitivity analyses, suggesting that the observed subgroup structure was not driven by a single modeling choice. Unsupervised learning identified symptom-based subgroups within concurrent constitutions and highlighted substantial internal heterogeneity. The findings suggest that clustering based on binary-encoded symptoms, combined with dimensionality reduction, may provide a useful framework for subgroup characterization and for future studies of stratified assessment and management.