A machine-learning model using routine clinical measures distinguished adults classified with phlegm-dampness constitution (PDC) from those with balanced constitution (BC) with very high performance in the study’s internal evaluation. The clinical model, called M1, was logistic regression. Its AUC, a measure of how well a model separates two groups, was 0.983, with a 95% confidence interval of 0.953 to 1.000. At a cut-off of 0.628, sensitivity was 0.945 and specificity was 0.970. The reported 95% confidence intervals for those measures ranged from 0.868 to 1.000 and from 0.913 to 1.000, respectively. Positive predictive value was 0.990, while negative predictive value was 0.900.
The data behind the result
The analysis included 139 Beijing residents aged 18 to 50: 96 PDC participants and 43 BC participants, classified under the 2009 traditional Chinese medicine Constitution standards. The outcome was a binary comparison of PDC with BC as the reference category. No included variable had missing values, so no participant was excluded for missingness and no imputation was required.
To build the models, the researchers screened clinical and biochemical variables with LASSO. From 17 screened variables, it retained age, waist circumference, body mass index and triglycerides as candidate predictors at a lambda value of 0.109. They developed logistic regression, random forest, support vector machine and XGBoost classifiers. The dataset was randomly divided into training and test sets at a 7:3 ratio, with stratification by PDC and BC. Standardization was fitted on the training data only, and cross-validation was used during model development.
A non-invasive candidate
The study also tested a non-invasive version, M2, based on systolic blood pressure, diastolic blood pressure and waist circumference. Its AUC comparison with M1 was not statistically significant, with the reported p-value greater than 0.05. The authors proposed M2 as a practical candidate. That finding is still an internal comparison, so the model needs testing in larger and more balanced populations before its performance can be judged more broadly.
What the model considered
To make the clinical model easier to interpret, the researchers used SHAP, a method for showing how inputs contribute to a model’s output. BMI and triglycerides were the key contributors to the clinical-model output, while fasting blood glucose and systolic blood pressure added complementary information. These are model-attribution findings, not evidence that those variables cause PDC. Their importance depends on the fitted model and the cohort analyzed.
Dietary signals shifted between datasets
Dietary preferences produced a different pattern across the two evaluation sets. XGBoost had the highest training AUC, 0.919, with a 95% confidence interval of 0.867 to 0.972 and specificity of 0.969. Logistic regression had the highest test AUC, 0.838, with a 95% confidence interval of 0.721 to 0.954 and specificity of 0.894. Its training AUC was 0.891, with a 95% confidence interval of 0.828 to 0.953, and its training specificity was 0.964.
In the dietary SHAP analysis, scorching and salty preferences shifted predictions toward PDC, while a light preference shifted predictions toward BC. The pattern describes the direction of the fitted model’s attributions. It does not establish that any dietary preference causes PDC.
Why the result needs wider testing
Several limits temper the headline numbers. The sample was relatively small and unevenly split between PDC and BC, and the selected indicators focused mainly on variables related to metabolic disease. The authors said those features may reduce model stability and generalizability and called for validation in larger, more balanced populations. Because the participants were recruited in Beijing and the models were evaluated in this dataset, the results should be read as evidence from the studied cohort, not as a validated rule for wider use.
Paper data and sources
Original title: An interpretable prediction model for non-invasive identification of the phlegm-dampness constitution based on machine learning
Authors: Zhang X, Li T, Liu Y et al.
Journal/Repository: Frontiers in endocrinology
Status: Peer-reviewed
First online: 2026-10-07
DOI: Not available
Original paper · Full text