JMIR Publications

ABSTRACT

Background:

Background:

Differentiating liver disease subtypes is challenging due to overlapping clinical features. While traditional mining methods rely on manual feature engineering, Large Language Models (LLMs) offer a new paradigm for integrating complex, multi-source clinical data.

Objective:

Objective:

To investigate the potential of LLMs in clinical data mining and evaluate their ability to support diagnostic reasoning across different granularities of liver disease classification.

Methods:

Methods:

This retrospective study analyzed data from 7,543 patients with liver disease (2010-2025). We utilized three LLMs (Qwen3, Huatuo-o1, II-Medical) to extract semantic embeddings from unstructured clinical narratives. To assess the model's depth of representation, we constructed a hierarchical classification framework: a three-class task (Autoimmune Liver Disease [AILD] vs. Drug-induced Liver Injury [DILI] vs. Chronic Hepatitis B [CHB]) followed by a more granular four-class task, where AILD includes the constituent subtypes autoimmune hepatitis (AIH) and primary biliary cholangitis (PBC). These LLM-derived features were fused with raw laboratory data to develop and compare machine learning classifiers.

Results:

Results:

The LLM-encoded framework demonstrated robust potential in both tasks, simplifying the modeling workflow by reducing manual variable selection. In the three-class task, the fusion models effectively captured the broad features of AILD. In the refined four-class task, the models maintained high discriminative power, proving that LLM embeddings can capture the subtle semantic nuances required to differentiate closely related subtypes like AIH and PBC. While direct zero-shot reasoning by LLMs showed exploratory value for diagnosis and medication, the embedding-based hybrid models provided more stable and accurate outcomes for complex multiclass identification.

Conclusions:

Conclusions:

LLMs possess significant potential for clinical data mining, offering a scalable strategy to integrate unstructured text with structured data. By successfully addressing both broad and fine-grained liver disease classification tasks, this framework provides a feasible path for developing intelligent decision support tools in real-world clinical settings.