1. Yulong Xu Henan University of Chinese Medicine
Three Data-Driven Approaches for Determining TCM Syndrome Subclasses in a Patient Population
Yulong Xu
Henan University of Chinese Medicine
Yanlong Li, Xiuyan Wu
Beijing University of Traditional Chinese Medicine
Nevin L. Zhang
The Hong Kong University of Science and Technology
This is joint work with Xu Yulong, Li Yanlong, and Wu Xiuyan.

2. Patient Classification is Important for TCM
TCM in practice
Divide patients into TCM Syndrome classes,
Apply different TCM treatments to different classes
Step 1 is called syndrome differentiation, or patient classification
The efficacy of TCM treatments depends heavily on whether the classification is done properly
Important also for
Randomized controlled trials to determine efficacy of treatment for a TCM syndrome
Biomedical basis for TCM syndromes
Our work is about patient classification. In TCM clinic practice, doctors first divide patients into TCM syndrome classes, and then apply different treatments to different classes. Here, the first step is called syndrome differentiation. We will simply called it patient classification. (Just classification, no explanation of mechanisms behind)
It goes with out saying that patient classification is very important. If it is done properly, your treatment will be more effective. If not, your treatment might not be effective, and might even make things worse for the patient.
Patient classification is also important for research work on TCM. For example, if you want to determine the efficacy of a treatment for a syndrome by conducting RCT, you need to first identify people with that syndrome. It is a patient classification problem. If you want to determine the biomedical basis for a TCM syndrome, you also need to identify people with that syndrome. Again, we have patient classification problem. In our view, patient classification is really a bottleneck of TCM research.

3. How Patients are Classified in Practice?
Judgment by individual doctors
Judgement by panel of experts’
Problem: Lack of consistency, subjective
In clinic practice, patient classification is done by individual doctors. It is known be subjective and lack consistency. Different doctors might give different diagnoses for the same patient. This is one of the biggest problems that faces TCM.
There are effort to set up standards for patient classification. This is done by panel of experts. This is better than individual doctors, but are still subjective and lack consistency. Different panels might recommend different standards.

4. Machine Learning: Supervised Learning
Illustrative example: Yang Deficiency
Labelled data:
Learn classification rule:
Problem: Class labels still lack consistency, subjective
Machine learning techniques have been used to study patient classification. One type of research is called supervised learning, where we start with a set of labelled data and try to learn a general classification rule.
In this illustrative example, we have three positive examples of yang deficiency and three negative examples of yang deficiency. From those six examples, we hope to learn a general classification rule. In particular, we wish to decide to which class this lady belongs. Obviously, she should be classified into the first class.
Like individual or panel judgments, supervised learning is also subjective and lacks consistency. The reason is that it needs the class labels, which are provided by individuals.

5. Machine Learning: Unsupervised Learning
No class labels in data
Classification rule: ?
In our work, the basic question we ask is this: Suppose we only have unlabeled data. That is, we have collected symptom inform of many people, but do not have diagnoses results. We do not class labels. Can we still establish patient classification rules?

6. Machine Learning: Unsupervised Learning
Cluster Analysis
Cluster interpretation: Top class is Yang Deficiency
Classification rule:
The answer is yes. The idea is to first perform cluster analysis and cluster the patient into several clusters. In this example, we can cluster those six people into two clusters.
Then we interpret those clusters. Suppose we think the cluster at the top correspond to Yang deficiency, and the cluster at the bottom is not Yang deficiency.
Then, suddenly we have labelled data. And we can then establish classification rules just as in the supervised case.

7. Machine Learning: Unsupervised Learning
Cluster Analysis
Three Approaches to cluster analysis
Latent class analysis
One-phase latent tree analysis
Two-phase latent tree analysis
Rest of talk: Brief explanation of the three methods w.r.t Qi Deficiency among VMCI Patient (VMCI = vascular mild cognitive impairment)
Now, the questions is: How do we do the cluster analysis? There are three methods: LCA, one-phase LTA and two-phase LTA.
This is why the work is entitled …Three Data-Driven Approaches for Determining TCM Syndrome Subclasses in a Patient Population
For the rest of the talk, I will briefly explain methods. I will do so by considering the task of determine Qi Deficiency among VMCI patient. That is, we have dataset of about 800 VMCI patients, and we want to determine which of the 800 people have Qi Deficiency.

8. Latent Class Analysis (LCA)
Symptoms related to Qi Deficiency manually selected based on domain knowledge
Patient partitioned into clusters using the following model,
Which says that whether the symptoms occur is totally determined by Qi Deficiency (Z), not any other factors.
4 clusters are obtained via model fitting and Expectation-Maximization EM
To perform latent class analysis, we first select from the data symptom variables that are related to Qi Deficiency. Those include, short of breath, chess oppression, palpitation, etc.
Then we create this model, which says that whether those symptoms occur is totally determined by whether the patient has Qi Deficiency and is not influenced by any other factors.
Then we try to divide the people into different number of clusters: 3 clusters, 4 clusters, 5 clusters, and so for. For each number, the actually partition is done using something called the EM algorithm. Then we choose the partition that fit the data the best.

9. Latent Class Analysis in Western Medicine Research
Van Smeden et al. Latent class models in diagnostic studies when there is no reference standard--a systematic review. Am J Epidemiol. 2014 Feb 15;179(4):423-31
Example: Subtypes of Major Depression
As it turns out, latent class analysis have been widely used Western Medicine Research. Because they sometimes need to divide a disease into subtypes even though is no gold standard. One example is to determine subtypes of major depression. This is very much like patient classification in TCM.
So, patient classification in the absence of gold standard is a problem faced not only by TCM, but also WM.
In WM research, Latent class analysis (LCA) has been widely applied to solve the problem. Naturally, it would be interesting to see whether LCA can be used to solve the syndrome classification problem of TCM.

10. Latent Tree Analysis (LTA)
In addition to class membership (Z), other factors also influence occurrence of symptoms
Y01 (便干便难，排便无力): Stool + Qi Deficiency
Y02 （纳呆，口淡）: Spleen + Qi Deficiency
W （心悸，胸闷，气短）: Heart + Qi Deficiency
In latent tree analysis, We use more complex models that look like this. It says that, the occurrence of symptoms are influenced not only by the class membership wrt Z, but also other factors.
Intuitively, this is more reasonable. For example, these two symptoms (dry stool or constipation) are both about stool. So we have a latent variable Y01 that represents the impact of qi deficiency on stool. These two symptoms (anorexia and bland taste in mouth) are both about spleen. So, we have a latent variables that represents spleen qi deficiency. Those last three symptoms (palpitation, chest opposition, and short of breath) are all related to the heart. So we have a latent variable W that represents heart qi deficiency. Because of those intermediate latent variables, this model is intuitively more reasonable than the latent class model.

11. How do we determine the intermediate latent variables?
One-phase latent tree analysis
symptoms related to qi deficiency + data =>
How do we determine the latent variables on the middle level. There are two methods.
One is one-phase LTA, which learns this three-level model direction from the part of data that contain only the symptoms related to qi deficiency

12. Two-Phase Latent Tree Analysis
Phase I: Analyze whole data to get global model
Phase II:
symptoms related to qi deficiency + global model =>
In two-phase LTA, we first analysis the entire data set, including symptoms not related to qi deficiency, and we get a global model. Then we obtain the tree-level model from the global model by looking only at the symptoms related to qi deficiency.

13. Objective Criteria Bayes Information Criterion (BIC)
Commonly used criterion for model selection in Statistics
Requires model to fit data well, but not to be overly complex
Two versions, same absolute value, different sign
Positive version： The smaller the better
Negative version: The larger the better
So, we have three methods. How do we evaluate them. In statistics, a commonly used criterion for model evaluation is Bayes Information Criterion or the BIC score. It requires a model to fit the data well, but not to be overly complex.
In this illustration, the dots are the data and the curves are the model. In the first figure, the model is too simple. It underfits data. In the last figure, the model fits the data well, but is to complex. The one in the middle is the best. It fits the data well and is not overly complex.
In the literature, BIC scores are sometimes given as positive values, and other times as negative values. They are equivalent. The absolute value is the same. For the positive version, the smaller the better, and for the negative version, the larger the better.

14. Comparisons on Cirrhosis Data: BIC Score
LTA is superior to LCA
Reason: LCA makes a unrealistic assumption called local independence. The assumption is relaxed in LTA.
One-phase LTA and two-phase LTA are similar
This table compare the three methods in terms of BIC scores. Here we use a data about cirrhosis. The task is to determine seven syndrome classes in the data, namely qi deficiency, qi stagnation, …..
We can see that the LTA methods are clearly better than LCA. The reason is that LCA makes an unrealistic assumption, and the assumption is relax.
The BIC scores for the two LTA methods are similar.

15. One-Phase LTA vs Two-Phase LTA
Model by two-phase LTA (top) more meaningful based on domain knowledge
Same in both models
口渴、口咽干燥：热证之口感反应
老舌、燥_糙苔： 热证之舌像反应
Two-phase LTA better
尿色深黄、便秘：热证于二便的表现
发热、数脉： 热证的整体表现
One-phase LTA is simpler
However, the model structures by the two-phase LTA are more meaningful than those by the one-phase LTA.
One-phase LTA fails in capture the patterns blue.
However, it is simpler to apply.

16. Results of Patient Clustering (by two-phase LTA)
Regarding Qi Deficiency in VMCI: Two clusters
Cluster 1 interpreted as Qi Deficiency.
Classification rule:
Qi Deficiency: Sore waist or knees (3.2), lack of strength (2.7), lassitude of limbs or body (2.7), short of breath (2.7), chest oppression (2.8), palpitation (2.8), etc
Threshold: 13.0
What do we get after patient clustering? Here is one example. The patients are divided into two clusters. 44% of the people are put into cluster 1. In this cluster, the symptom sore waist or knew, lacks of strength, etc occur with high probabilities. In the other cluster, they occur with low probabilities. Hence, we interpret cluster 1 as Qi deficiency.
Then, using the information from the table, we can establish a classification rule for Qi Deficiency. Here, each symptom is given a score. If the total score exceed the threshold, we can conclude that the patient is in the Qi Deficiency class.

17. Summary
LTA is superior to LCA in terms of both how well the underlining probabilistic models fit data and how well the model fit TCM theories.
Between the two variations of LTA, the two-phase method is better in terms of fit to TCM theories
while the one-phase method is more convenient to use in practice.
To summarize, ..

18. All Three Approaches Supported by Lantern Software
Finally, all the three methods are supported in a softw are called Lantern that we have develop. It is online and free.