Diagnostics · Nov 14, 2026
Enhancing HCC Surveillance with GPTs: Computational, Prompt Engineering, and Multimodal Strategies
By Wu SH, Lin XX, Zeng Y et al.
Objectives: Research on applying Large Language Models to hepatocellular carcinoma (HCC) surveillance is limited. We aimed to develop customized GPTs for HCC surveillance and evaluate their performance in predicting HCC risk and providing surveillance recommendations. Materials and Methods: This single-center retrospective study included 376 patients with 797 surveillance visits between January 2014 and March 2023 for evaluation. We developed GPTs by integrating medical expert knowledge and optimized it through three strategies: (1) boosting computational capabilities via code enhancements and in-context learning; (2) enhancing reasoning abilities through prompt engineering; (3) increasing clinical applicability by integrating multimodal clinical data. We assessed the predictive performance of GPTs across different modalities and the efficiency and clinical appropriateness of their surveillance recommendations. Results: Combining in-context learning with code instruction improved HCC risk score calculation accuracy to 97.4%, compared with 90.2% using code instruction alone and 31.1% using default calculation (both p < 0.001). For the five-year HCC prediction, the second chain-of-thought (CoT2) strategy improved sensitivity from 65.3% to 77.6% (p = 0.037), with a numerically higher AUC compared with original GPTs (0.782 vs. 0.731, all p = 0.095). Dual-modality and multi-modality GPTs outperformed traditional scores in predicting HCC (AUC: 0.839 and 0.840 vs. 0.737–0.762, all p < 0.05). Among patients who developed HCC, multi-modality GPTs achieved the highest surveillance efficiency index (2.34) and an estimated earlier detection time of 5.43 months, compared with 2.22 and 4.49 months, respectively, under AASLD-based surveillance, with a higher estimated annual CT/MRI examination frequency (1.46 vs. 0.98). Conclusions: Customized GPTs showed promise for HCC risk prediction and individualized surveillance, warranting further prospective multicenter validation.