Introduction
Traditional Chinese Medicine (TCM) offers a unique perspective on disease management through "syndrome differentiation." Unlike Western medicine, which often focuses on a specific pathogen, TCM analyzes a patient's entire clinical presentation—including pulse, tongue appearance, and subjective symptoms—to categorize their functional state into a "syndrome." For complex conditions like AIDS, this diagnostic process is vital for personalized treatment but remains highly subjective and difficult to standardize.
The Challenge of TCM Syndrome Differentiation
In AIDS patients, symptoms are often diverse and overlapping. A single patient might present with dozens of clinical markers, many of which are "background noise"—common symptoms that do not help distinguish one syndrome from another. Traditional machine learning models, such as Support Vector Machines (SVM) or k-Nearest Neighbors (k-NN), often struggle because they treat all input features with equal weight. This lack of prioritization leads to lower diagnostic accuracy and a "black box" problem where the reasoning behind a diagnosis is opaque.
The ATT-MLP Framework
To address these challenges, researchers developed the ATT-MLP (Attention-based Multilayer Perceptron) framework. The core innovation is an integrated Attention Module that precedes the classification stage.
As shown in the architecture above, the model does not process symptoms blindly. Instead:
- Attention Weighting: The system calculates a specific weight for each symptom using a non-linear activation function. This mimics the expertise of a senior TCM practitioner who knows which specific signs (e.g., a "string-taut pulse") are diagnostic "signals" versus general "noise" (e.g., "fatigue").
- Feature Refinement: The raw symptom vector is multiplied by these weights, effectively amplifying critical diagnostic markers and suppressing irrelevant ones.
- Classification: The weighted features are then processed by a three-layer MLP to predict the most likely syndrome among seven predefined categories.
Experimental Setup and Data
The model was trained and validated using a large-scale, real-world dataset from a TCM pilot project in China.
- Data Scale: 10,910 clinical records collected between 2004 and 2013.
- Features: 93 distinct clinical symptoms were used as inputs.
- Labels: Seven specific TCM syndromes (S1 through S7) served as the target classes.
- Validation: 10-fold cross-validation was employed to ensure the robustness of the results.
Results and Interpretability
The ATT-MLP model demonstrated significant performance gains over baseline methods. It achieved an average accuracy of 85.7%, a substantial improvement over Naive Bayes (74.8%), SVM (72.3%), and a standard MLP without attention (76.0%). Furthermore, the model showed high specificity (91.3%), indicating a strong ability to avoid false-positive diagnoses.
Beyond raw performance, the attention mechanism provides critical interpretability. By visualizing the attention weights, researchers can confirm that the model's "focus" aligns with TCM theory.
The heatmap reveals how the model prioritizes different symptoms for different syndromes. For instance, specific tongue coatings or pulse types receive high weights for the syndromes they characteristically define, while generalized symptoms like "cough" receive lower weights across the board. This transparency bridges the gap between modern AI and ancient diagnostic wisdom.
Conclusion
The ATT-MLP framework represents a significant step toward objective, automated TCM diagnostics. By intelligently filtering clinical noise through an attention mechanism, the model achieves superior accuracy while remaining interpretable to practitioners. Future work will focus on evolving this system into a multi-label classification framework to better reflect the reality of patients suffering from multiple overlapping syndromes simultaneously.
Summary updated