A multimodal deep learning model predicting hyperprogressive disease for PD-1 blockade in advanced hepatocellular...

Area of Science:

• Oncology
• Artificial Intelligence
• Medical Imaging

Background:

• Immune checkpoint inhibitors targeting programmed cell death 1 (PD-1) are used for advanced hepatocellular carcinoma (Ad-HCC).
• Treatment responses are heterogeneous, with hyperprogressive disease (HPD) being a significant concern.
• Lack of reliable pre-treatment tools to identify patients at high risk for HPD.

Purpose of the Study:

• To develop and validate a multimodal AI model for predicting HPD in Ad-HCC patients receiving PD-1 inhibitor-based triple therapy.
• To assess the model's performance against clinical-only and imaging-only approaches.
• To establish a clinically interpretable decision-support tool for HPD risk stratification.

Main Methods:

• Multicenter retrospective study of 665 patients with Ad-HCC.
• Development of a transformer-based multimodal model (HOPE) integrating computed tomography (CT) imaging (arterial and portal phases) with clinical factors.
• Internal and external validation of the HOPE model.

Main Results:

• HOPE achieved an area under the receiver operating characteristic curve (AUC) of 0.801 in internal validation and 0.687 in external validation.
• The multimodal model outperformed clinical-only and imaging-only baseline models.
• Ablation analyses confirmed the value of multimodal integration, supported by subgroup analyses and survival risk stratification.

Conclusions:

• The HOPE model demonstrates potential as a clinically interpretable pre-treatment decision-support tool for HPD risk stratification in Ad-HCC.
• HOPE may aid in closer monitoring and risk-adapted management strategies for high-risk patients.
• Multimodal AI integration shows promise for improving outcomes in Ad-HCC treated with PD-1 inhibitors.