LLM-BCgrading: Large language model-based Chinese medical long text classification for bladder cancer grade...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Oncology
• Natural Language Processing

Background:

• Traditional bladder cancer (BC) grading relies on invasive cystoscopic biopsies.
• These methods are prone to sampling errors and interobserver variability.
• There is a need for noninvasive, accurate BC grading techniques.

Purpose of the Study:

• To develop and evaluate a large language model (LLM)-based noninvasive approach for predicting bladder cancer (BC) grade.
• To utilize long Chinese medical texts, including admission records and CT urography (CTU) descriptions, for BC grade prediction.
• To enhance the model's performance using a gated multiplicative attention mechanism (GMAM) and cost-sensitive optimization.

Main Methods:

• Retrospective collection of admission records and CTU descriptions from 642 BC patients.
• Development of LLM-BCgrading leveraging HuatuoGPT-7B for Chinese medical text representation.
• Integration of GMAM for feature emphasis and optimization with a cost-sensitive loss function to address class imbalance.

Main Results:

• The best model, combining admission records and CTU descriptions with GMAM, achieved a balanced accuracy of 0.757, macro F1 score of 0.749, and macro AUC of 0.740.
• Attention-based fusion of both text types significantly improved performance.
• GMAM outperformed conventional attention mechanisms, and 256 was identified as the optimal embedding size.

Conclusions:

• LLMs can accurately predict bladder cancer grade noninvasively using Chinese medical long-texts.
• The developed LLM-driven framework demonstrates robustness and clinical relevance through attention-based fusion and cost-sensitive optimization.
• Shapley additive explanations support the interpretability and clinical utility of the model.