Digital health · 2025
LLM-BCgrading: Large language model-based Chinese medical long text classification for bladder cancer grade prediction.
СтатьяНазвание и аннотация: Перевод пока не подготовлен
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Pan Xianwei, Wen Lijie, Li Yuhua, Zhang Yijia, Lu Mingyu
Аннотация
BACKGROUND: Traditional cystoscopic biopsy-based methods for histological grading of bladder cancer (BC) are invasive, subject to sampling errors, and susceptible to interobserver variability among pathologists. To address these challenges, this study explores a large language model (LLM)-based noninvasive approach to BC grade prediction using long Chinese medical texts.
METHODS: We retrospectively collected admission records and computed tomography urography (CTU) descriptions from 642 patients pathologically diagnosed with BC. Each paired text was annotated as low grade or high grade according to histopathological results. We developed LLM-BCgrading to leverage HuatuoGPT-7B for Chinese medical long-text representation and integrated a gated multiplicative attention mechanism (GMAM) to selectively emphasize discriminative features. To address class imbalance and clinical risk asymmetry, the model was optimized with a cost-sensitive loss function. Performance was evaluated on a fixed internal test set with additional evaluation on an independent external validation cohort to assess generalizability.
RESULTS: The best-performing configuration combined both admission records and CTU descriptions via an attention-based fusion strategy and GMAM, achieving balanced accuracy of 0.757, macro F1 score of 0.749, and macro AUC of 0.740. The ablation results demonstrated that incorporating both texts significantly improved classification performance compared with single-text configurations, and the GMAM consistently outperformed conventional attention mechanisms. Dimensionality experiments identified 256 as the optimal embedding size, balancing computational efficiency and predictive performance.
CONCLUSION: Our findings demonstrate that LLMs can effectively process Chinese medical long-texts for accurate preoperative prediction of BC grade. Attention-based fusion, cost-sensitive optimization, and interpretability based on Shapley additive explanations further support the robustness and clinical relevance of this LLM-driven framework.
METHODS: We retrospectively collected admission records and computed tomography urography (CTU) descriptions from 642 patients pathologically diagnosed with BC. Each paired text was annotated as low grade or high grade according to histopathological results. We developed LLM-BCgrading to leverage HuatuoGPT-7B for Chinese medical long-text representation and integrated a gated multiplicative attention mechanism (GMAM) to selectively emphasize discriminative features. To address class imbalance and clinical risk asymmetry, the model was optimized with a cost-sensitive loss function. Performance was evaluated on a fixed internal test set with additional evaluation on an independent external validation cohort to assess generalizability.
RESULTS: The best-performing configuration combined both admission records and CTU descriptions via an attention-based fusion strategy and GMAM, achieving balanced accuracy of 0.757, macro F1 score of 0.749, and macro AUC of 0.740. The ablation results demonstrated that incorporating both texts significantly improved classification performance compared with single-text configurations, and the GMAM consistently outperformed conventional attention mechanisms. Dimensionality experiments identified 256 as the optimal embedding size, balancing computational efficiency and predictive performance.
CONCLUSION: Our findings demonstrate that LLMs can effectively process Chinese medical long-texts for accurate preoperative prediction of BC grade. Attention-based fusion, cost-sensitive optimization, and interpretability based on Shapley additive explanations further support the robustness and clinical relevance of this LLM-driven framework.