Abstract
Traditional Chinese medicine (TCM) is one of the safe and effective methods to treat liver cirrhosis. The practitioners of TCM assess hepatic function in term of syndrome. But the course of syndrome differentiation is subjectivity. At present most of all the researches are focused on the relationship between the syndrome and the Western medicine objective indicators such as child-pugh grade. In fact syndrome is the synthesis of signs and symptoms and collecting signs, symptoms is easy than syndrome differentiation. We try to explore the relationship between the objective Western medicine standard such as child-pugh grade, decompensation or compensation stage, active or inactive period and the signs and symptoms of TCM by using the data mining method. We use the information gain method to assess the attributes and use five typical classifiers such as logistic, Bayes-Net, NaiveBayes, RBF and C4.5 to obtain the classification accuracy. After attribute selection, we obtain the main symptoms and signs of TCM relating to the stage, period and child-pugh grade about liver cirrhosis. The experiment results show the classification accuracy is improved after filtering some symptoms and signs.
Introduction
In treating liver cirrhosis, Traditional Chinese Medicine (TCM) has shown clinical effectiveness, but its diagnostic process is inherently subjective. TCM practitioners rely on "syndrome differentiation"—a holistic assessment based on inspection, smelling, inquiry, and palpation. Western medicine, conversely, relies on objective lab parameters and standardized scoring systems like the Child-Pugh grade.
Previous research has attempted to link TCM syndromes to Western indicators. However, because determining the "syndrome" requires subjective expert judgment, this approach is difficult to standardize. This study addresses this bottleneck by bypassing the subjective "syndrome" step entirely, using machine learning to map raw, easily observable TCM symptoms and signs directly to objective Western medical indicators.
Methodology
The cross-disciplinary mapping is treated as a supervised classification problem using a standard machine learning pipeline:
- Attribute Selection: Information Gain (based on entropy) is used to evaluate the predictive power of each TCM symptom/sign with respect to the Western medical label.
- Discretization: Continuous numeric variables, such as the avoirdupois exponent, are converted into discrete bins using an entropy-based method governed by the Minimum Description Length (MDL) principle.
- Classification: Five standard classifiers are evaluated: Logistic Regression, BayesNet, NaiveBayes, Radial Basis Function (RBF) networks, and C4.5 (decision trees).
- Validation: Models are evaluated using 10 iterations of 10-fold cross-validation to ensure robustness on the clinical dataset.
Datasets
The study utilizes a clinical dataset of 294 patient cases provided by a liver cirrhosis expert from the Shanghai University of Traditional Chinese Medicine. Features include 40 TCM symptoms (scored 1-4), 27 TCM signs (scored 1-2), age, sex, and an avoirdupois exponent. These are evaluated against three Western medical targets:
- Stage: Compensation (111 cases) vs. Decompensation (183 cases).
- Child-Pugh Grade: Grade A (109 cases), B (93 cases), or C (92 cases).
- Period: Active (212 cases) vs. Inactive (82 cases).
Results
Filtering out irrelevant attributes yielded significant improvements in classification accuracy across all targets. For instance, accuracy for the Compensation vs. Decompensation stage jumped from a baseline of 58.8% to 80.9% using Logistic Regression.
Logistic Regression proved best for binary predictions (stage and period), while the C4.5 decision tree performed best for the multi-class Child-Pugh grade prediction. Notably, Logistic Regression was highly sensitive to the number of attributes, whereas Bayesian methods remained more robust to feature bloat.
Discussion
The results demonstrate that "less is more" in medical machine learning. Filtering out "weak" attributes using Information Gain is crucial, as many traditional symptoms are redundant for predicting specific Western clinical stages. Furthermore, the study identifies specific symptom clusters that bridge TCM observations and Western diagnostic criteria.
As shown in Table 2, abdominal distension, yellow eyes/urine, and puffy lower limbs were identified as the strongest predictors for the Western "decompensation" stage. This mapping provides a more objective basis for TCM assessment in clinical settings.
Conclusion
While classification accuracy improved significantly, peaking around 70-80%, it is not yet high enough for standalone clinical reliance. The small, single-institution dataset limits generalizability. Future work will focus on larger, multi-center datasets and advanced modeling techniques to capture complex symptom interactions and further objectify TCM theory.
Summary updated