AI-powered TCM Diagnosis: A Multi-Task Learning Approach for Syndrome Element Diagnosis and Syndrome Differentiation

Background Syndrome differentiation is the cornerstone of Traditional Chinese Medicine (TCM). The development of robust artificial intelligence aided diagnosis (AIAD) syndrome differentiation method is therefore a pivotal direction for TCM modernization. Contemporary TCM theory identifies syndrome elements as fundamental components essential for accurate syndrome differentiation. However, existing computational methods have not fully exploited their potential through joint learning approaches. This study proposes a novel AIAD framework that simultaneously learns syndrome element diagnosis and syndrome differentiation. The framework aims to enhance diagnostic accuracy and provide empirical validation of the theoretical role of syndrome elements within the TCM diagnostic process. Methods We developed a novel AIAD framework trained and tested on 6,226 electronic health records (EHRs) and further evaluated on an independent dataset of 1,057 EHRs derived from clinical guidelines. A multi-task learning approach was utilized to simultaneously model syndrome element diagnosis and syndrome differentiation. We proposed and compared two models: AIAD-A1, which considers syndrome elements as intermediate variables, and AIAD-A2, which treats them as mediators. Model performance was assessed using top-K accuracy and NDCG@K metrics. Results . The AIAD-A2 model, which treats syndrome elements as mediators, significantly outperformed both the AIAD-A1 model and baseline models across multiple evaluation metrics. AIAD-A2 improved top-K accuracy by up to 6.31% and demonstrated remarkable enhancements in handling long-tail data, improving top-15 accuracy by 21.34% within the tail subgroup. Additionally, it exhibited superior generalizability, with the smallest performance decline (7.77% in top-15 accuracy) on an independent test dataset facing a different data distribution, compared to drops of 22.7% for the baseline and 12.44% for AIAD-A1. Moreover, model visualization confirmed AIAD-A2’s capability to focus on the most clinically relevant symptoms. Conclusions Our findings reveal the role of syndrome elements as mediators in TCM syndrome differentiation. Accurate diagnosis necessitates the integrated consideration of both syndrome elements and the original symptom descriptions. The proposed AIAD-A2 framework offers an effective and generalizable approach for AI-powered TCM diagnosis, successfully addressing challenges such as data imbalance and distribution shifts. This work contributes to the modernization of TCM by delivering a robust AI method and deepening the theoretical understanding of the diagnostic process.