Model Aids IPF Diagnosis and Treatment in TCM

A new study offers practitioners of traditional Chinese medicine (TCM) a machine learning-based approach to diagnosing and treating idiopathic pulmonary fibrosis (IPF). The report was published in the journal Scientific Reports.¹

TCM is built on a process called syndrome differentiation, a method of classifying different diseases using a traditional methodological system, explained first author Hua Ye, PhD, of the Chengdu University of Traditional Chinese Medicine, and colleagues. However, the method of obtaining and making sense of medical information in TCM is quite different from Western medicine, Ye and colleagues said.

“TCM data is characterized by nonlinearity, ambiguity, unstructuredness, and multidimensionality,” they said.

As artificial intelligence (AI) and machine learning have become more and more common in healthcare, a growing body of research has been devoted to figuring out how to apply the new technologies to TCM, the authors said.

Such questions are particularly important for life-threatening respiratory diseases like IPF, a disease that is easy to misdiagnose and which follows an unpredictable course.²

Previous research has suggested that a type of neural network called a back propagation (BP) neural network is well-suited to TCM, but Ye and colleagues said it is not without limitations.¹

“These drawbacks can significantly affect the accuracy of the model classification,” the investigators said

However, Ye and colleagues said there are ways to modify BP neural networks to make them a better fit for TCM. In the new study, the investigators analyzed a hybrid strategy that combines two optimization techniques—the Levenberg Marquardt (LM) algorithm and the genetic algorithm (GA). When applied to the BP neural network, these algorithms helped to enhance the convergence and generalization performance of the network. In other words, the authors believed the hybrid algorithm would make the BP neural network faster and more useful in the real world. If the approach worked, it would make it easier for TCM practitioners to accurately diagnose a patient’s syndrome and prescribe the right therapy to treat IPF.

Another issue with machine learning is that too many redundant or irrelevant features in patient data can negatively interfere with the model’s performance. Ye and colleagues said feature screening using a method called mean impact value (MIV) can help remove unnecessary features and improve classification accuracy.

“TCM syndrome differentiation focuses on individual differences and dynamic changes, and contains a large amount of complex information, and the selection of specific symptoms related to the diagnosis is crucial for the establishment of a syndrome classification model,” they said.

The authors called their hybrid model the MGLB model, which is short for MIV, GA, the LM algorithm, and BP neural networks. The model was developed using 956 real-world cases of IPF patients who sought care via TCM.

The investigators compared their model to 3 other existing types of analysis and found that their model had the highest classification accuracy, at 81.22%, which they said represented an improvement both on accuracy and stability.

“More importantly, the MIV-based feature screening enables transparent mapping between symptoms and syndromes, aligning well with TCM diagnostic logic,” Ye and colleagues said.

The investigators said their system helps curb some of the limitations of BP neural networks and creates a bridge between TCM and ML. They said their model could also act as an effective auxiliary means of TCM syndrome differentiation to help less experienced TCM practitioners and improve the shortcomings of the traditionally subjective form of medical care.

“In summary, this study presents a novel and interpretable machine learning-based framework for TCM syndrome classification, integrating feature screening, hybrid optimization, and real-world data validation to support accurate and interpretable diagnosis of IPF,” the authors concluded.

References

1. Ye H, Gong W, Yuan P, Zhang R, He B, Lin W. A supported decision-making model for idiopathic pulmonary fibrosis based on feature screening and optimized neural network. Sci Rep. Published online November 28, 2025. doi:10.1038/s41598-025-10070-6

2. Mai TH, Han LW, Hsu JC, Kamath N, Pan L. Idiopathic pulmonary fibrosis therapy development: a clinical pharmacology perspective. Ther Adv Respir Dis. 2023;17:17534666231181537. doi:10.1177/17534666231181537