AI Framework Halves Symptom List for Traditional Chinese Medicine Syndrome Diagnosis

Researchers have built a two-stage artificial intelligence framework that uses a large language model to extract symptoms from Traditional Chinese Medicine (TCM) records and then applies evolutionary algorithms to trim the resulting feature set, cutting the candidate symptom list roughly in half without a statistically detectable loss in classification performance. The framework, called LEAF (LLM extraction and multiobjective feature selection), was evaluated on 3,253 patient records covering 14 respiratory syndromes from the TCM-SD benchmark.

The motivation is interpretability. Differentiating TCM respiratory syndromes requires weighing overlapping symptoms together with tongue and pulse findings, and most existing AI systems for TCM perform this task end-to-end, without an interpretable intermediate representation. LEAF addresses that gap by first extracting a 35-item symptom ontology, built entirely from public TCM sources, and then selecting features over that ontology rather than operating on raw text.

In the extraction stage, the team compared three interchangeable large language models. Qwen2.5-14B-Instruct ranked best, ahead of Phi-4 and the TCM domain model BianCang. The ontology these models produced serves as the input representation for the second stage of the pipeline.

That second stage benchmarked four multiobjective evolutionary algorithms: the non-dominated sorting genetic algorithm II (NSGA-II), the strength Pareto evolutionary algorithm 2 (SPEA2), a competitive mechanism-based multiobjective particle swarm optimizer (CMOPSO), and non-dominated sorting differential evolution (NSDE). Each algorithm was run over ten seeds, and the results were compared using Friedman’s and Nemenyi’s statistical tests on both classification outcome and search-process quality.

On held-out macro-F1, the four algorithms showed no significant separation, although NSDE led nominally, and none of them exceeded the full-feature baseline on point estimates. In contrast, hypervolume, a measure of search-process quality, differed significantly among the algorithms and favored NSGA-II, which was actually the weakest of the four on held-out macro-F1. Inverted generational distance behaved similarly as a proxy. The authors conclude that hypervolume and inverted generational distance are not reliable stand-ins for classification-relevant algorithm selection in this setting.

A four-classifier panel reinforced the picture of statistical equivalence. No algorithm was significantly different from either the full-feature baseline or a random-feature control, and only the full-feature baseline significantly outperformed the random-feature control. In other words, the evolutionary selection achieved its roughly 50 percent reduction in the candidate feature set at a small point-estimate cost in performance, but that cost was not statistically detectable.

Applying NSDE to the remaining two LLM pipelines reproduced the original model ranking, suggesting the findings are not tied to a single extraction model. The representative solution produced by NSDE was fully attested against the TCM-SD knowledge base, meaning every symptom in the selected set could be verified against that source, although this full attestation did not typically hold across all ten seeds.

The study also demonstrated a practical alternative to human annotation: a no-gold-standard proxy evaluation can substitute for expert labeling when expert time is limited. This matters because expert annotation of TCM syndromes is a bottleneck for building and validating such systems.

Because both stages of LEAF are modular and plug in and out freely, the framework can absorb stronger components in either stage as they become available, such as improved language models or better evolutionary optimizers. The work was published in Frontiers in Medicine as an original research article by Xuanhang Luo, Huan Ding, Xize Zhang, Binhuan Miao and Li Shi, affiliated with Changchun University of Chinese Medicine and its affiliated hospital.

The results suggest that an LLM-extraction front end built exclusively from public knowledge can supply a representation over which evolutionary feature selection operates effectively, offering an interpretable intermediate step for TCM respiratory syndrome differentiation while keeping classification performance essentially intact.

Subject of Research: A modular AI framework combining large language model symptom extraction with multiobjective evolutionary feature selection for interpretable Traditional Chinese Medicine respiratory syndrome differentiation.

Article Title: LEAF: benchmarking multiobjective evolutionary feature selection over large language model–extracted symptoms for interpretable Traditional Chinese Medicine respiratory syndrome differentiation

Article References: Luo, X., Ding, H., Zhang, X., Miao, B., & Shi, L. (2026). LEAF: benchmarking multiobjective evolutionary feature selection over large language model–extracted symptoms for interpretable Traditional Chinese Medicine respiratory syndrome differentiation. Frontiers in Medicine, 13, 1990580. https://doi.org/10.3389/fmed.2026.1990580

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.1990580

Keywords: Traditional Chinese Medicine, large language models, evolutionary computation, multiobjective feature selection, syndrome differentiation, respiratory diseases, macro-F1, hypervolume, NSGA-II, NSDE, interpretability, TCM-SD benchmark