Enhancing hepatocellular carcinoma diagnosis in non-high-risk patients: a customized ChatGPT model integrating contrast-enhanced ultrasound.

Abstract

Purpose: This study aims to improve hepatocellular carcinoma (HCC) diagnostic accuracy in non-high-risk populations by utilizing GPTs that incorporate integrated risk coefficients, and to explore its feasibility.

Material and methods: Between August 2016 and June 2019, patients with focal liver lesions (FLLs) in non-high-risk populations, confirmed by histopathology or clinical/imaging evidence, were retrospectively included. A logistic regression model was developed using baseline characteristics and contrast-enhanced ultrasound (CEUS) features to identify independent HCC risk factors. Three ChatGPT-based models were evaluated: ChatGPT 4o (a general-purpose model developed by OpenAI), BaseGPT (a customized model with HCC diagnostic knowledge), and RiskGPT (a further customized model integrating HCC knowledge and identified risk factors). Their intra-agreement and diagnostic performance were compared.

Results: Logistic regression identified male, obesity, HBcAb or HBeAb positivity, elevated alpha-fetoprotein, and mild washout on CEUS as associated with HCC. RiskGPT achieved the highest area under a receiver operating characteristic curve (AUC) (0.89) and demonstrated superior accuracy (90.3%) in HCC identification; significantly outperforming both ChatGPT 4o (AUC 0.79, P = 0.002; accuracy 83.1%, P = 0.02) and BaseGPT (AUC 0.81, P = 0.008; accuracy 80.6%, P = 0.002). RiskGPT demonstrated superior sensitivity compared to ChatGPT 4o (85.5% vs. 66.3%) and outperformed BaseGPT in specificity (92.7% vs. 80.6%) and positive predictive value (85.5% vs. 67.7%) (all P < 0.001). Additionally, RiskGPT showed substantial intra-consistency in diagnosing FLLs, with a κ value of 0.78.

Conclusion: RiskGPT improves HCC diagnostic accuracy in non-high-risk patients by integrating clinical, imaging features, and risk coefficients, demonstrating significant diagnostic potential.