Abstract
Tibetan medicine, a traditional ethnic medical system with a deep history in China, plays a vital role in the healthcare of the northwestern plateau regions. Despite its clinical efficacy, the subjective nature of its diagnostic system presents challenges for modern standardization. This paper focuses on Chronic Atrophic Gastritis (CAG), a high-incidence gastrointestinal disease in plateau areas. We propose an improved back-propagation (BP) neural network model for the classification and prediction of Tibetan medical syndromes. Our methodology integrates K-means clustering for objective label discovery and Correlation-based Feature Selection (CFS) for dimensionality reduction. By optimizing the BP network with Gaussian weight initialization, exponential learning rate decay, and L2 regularization, we achieved a classification accuracy of 99.09%, significantly outperforming standard models.
Introduction
Tibetan medicine has evolved over 3,800 years, developing a unique diagnostic framework that includes pulse, urinary, and tongue examinations. However, unlike Western medicine or even Traditional Chinese Medicine (TCM), it lacks a robust, data-driven standardization for syndrome classification. Chronic Atrophic Gastritis (CAG) is a particularly relevant case study due to its high prevalence in the plateau regions of China. Currently, CAG diagnosis in Tibetan medicine relies heavily on the subjective experience of practitioners, which can lead to inconsistencies.
The application of data mining in this field faces two primary hurdles: the lack of objective "ground truth" labels in clinical datasets and the inherent limitations of standard neural networks, such as slow convergence and a tendency to overfit on small, high-dimensional medical datasets. This research addresses these challenges by developing a multi-stage pipeline to transform raw clinical observations into a high-precision predictive tool.
Methodology
The research utilized a clinical dataset of 223 CAG patients from the Qinghai Provincial Tibetan Hospital. The raw data comprised 81 diagnostic attributes, including 18 pulse types, 16 urinary exam features, 7 tongue diagnosis indicators, and 20 general symptoms.
Unsupervised Label Discovery and Feature Selection
To overcome the lack of pre-existing syndrome labels, the authors employed K-means clustering. By applying the "elbow rule" and analyzing the average contour coefficient (silhouette score), the researchers determined that the patient data naturally clustered into two distinct syndrome types. These clusters were later validated against the expert opinions of professional physicians, showing a high degree of alignment (up to 78.6%).
To streamline the model and avoid the "curse of dimensionality," Correlation-based Feature Selection (CFS) was used. This process reduced the 81 diagnostic items to the 9 most impactful features, including specific indicators of atrophy gland reduction and intestinal metaplasia.
The Improved BP Neural Network
The core of the predictive model is a Back-Propagation (BP) neural network, optimized to handle the specific constraints of medical data mining.
The network architecture, illustrated above, consists of an input layer for the selected features, a hidden layer for processing, and an output layer for syndrome prediction. To improve performance, three key modifications were introduced:
- Gaussian Weight Initialization: Instead of random initialization, weights were set using a Gaussian distribution scaled by the square root of the input size. This prevents neurons from saturating early in the training process.
- Exponential Decay Learning Rate: The model utilizes a dynamic learning rate that starts high to ensure rapid initial convergence and decreases exponentially to allow for precise weight adjustments as the model nears the global optimum.
- L2 Regularization: To combat overfitting—a common issue with small datasets—a penalty term was added to the loss function, discouraging the model from becoming overly complex.
As shown in the neuron model above, each unit calculates a weighted sum of its inputs plus a bias, which is then passed through an activation function $f(\cdot)$ to produce an output.
Results and Discussion
The improved BP network was tested against several standard classification algorithms. The results highlight the effectiveness of the proposed optimizations.
| Model | Accuracy |
|---|---|
| J48 Decision Tree | 96.80% |
| SMO (SVM) | 97.26% |
| Standard MLP | 98.64% |
| Improved BP Network | 99.09% |
The optimized model reached convergence in just 500 iterations, significantly faster than standard implementations. Furthermore, the model's computational efficiency is noteworthy, with a total time for building and prediction of 0.931 seconds. This speed makes the system viable for real-time clinical auxiliary decision support, assisting doctors in making more objective and standardized diagnoses.
Conclusion
This study presents a successful framework for the standardization and prediction of Tibetan medical syndromes. By combining unsupervised clustering with an optimized neural network, the researchers demonstrated that machine learning can objectively mirror and even enhance traditional medical intuition. While the current sample size is a limitation, the high accuracy achieved for CAG provides a strong foundation for future work. Expanding this model to other plateau diseases and larger patient cohorts will be essential for integrating these digital health tools into the medical infrastructure of northwestern China.
Summary updated