Clinician-deployable deep hypergraph model integrating clinical and CT radiomics predicts immunotherapy outcomes in NSCLC.
Authors
Affiliations (13)
Affiliations (13)
- School of Health Management, China Medical University, Shenyang, Liaoning, China.
- Department of Pulmonary Surgery, Guangdong Lung Cancer Institute, Guangdong Provincial People's Hospital (Guangdong Academy of Medical Sciences), Southern Medical University, Guangzhou, China.
- Ganzhou Municipal Hospital, Ganzhou, Jiangxi, China.
- Department of Laboratory Medicine, Nanfang Hospital, Southern Medical University, Guangzhou, China.
- Faculty of Data Science, City University of Macau, MacauChina.
- Department of Radiology, Shengjing hospital of China Medical University, Shenyang, Liaoning, China.
- Department of Medical Oncology, Guangdong Provincial People's Hospital, the First Hospital of China Medical University, Shenyang, Liaoning, China.
- Shengjing Hospital of China Medical University, Shenyang, Liaoning, China.
- Department of Oncology, Guangdong Provincial People's Hospital (Guangdong Academy of Medical Sciences), Southern Medical University, Guangzhou, China.
- Department of General Practice, The First Affiliated Hospital of Guangzhou Medical University, Guangzhou, China.
- Department of Radiology, The First Affiliated Hospital of University of Science and Technology of China (USTC), Division of Life Sciences and Medicine, USTC, Hefei, Anhui, China.
- Department of Thoracic Surgery, Shengjing hospital of China Medical University, Shenyang, Liaoning, China.
- Department of Radiology, Guangdong Provincial People's Hospital (Guangdong Academy of Medical Sciences), Southern Medical University, Guangzhou, China.
Abstract
Current image-based deep learning models that predict the benefits of immunotherapy in non-small cell lung cancer (NSCLC) require high-performance hardware. We aimed to develop and externally validate a clinician-operable prognostic model that integrates clinical and imaging data in a format usable by clinicians on standard central-processing-unit (CPU) hardware. This multicenter study included 1,379 patients with NSCLC treated with immunotherapy from five Chinese hospitals and the Memorial Sloan-Kettering (MSK) Cancer Center. A pairwise association encoder (PAE) converted routinely collected baseline clinical variables into edge weights of a patient-similarity graph, while radiomics features extracted from pretreatment chest computed tomography (CT) were embedded as node attributes in a Deep Hypergraph for NSCLC (DHGN). DHGN was trained on progression-free survival (PFS) and validated on overall survival (OS). Model performance was compared with three established control models and a published deep learning benchmark. Ten thoracic oncology experts independently implemented and tested each pipeline. Trained on PFS, DHGN yielded C-indices of 0.72, 0.71, and 0.71, and 0.70, 0.71, and 0.69 when validated on OS. DHGN outperformed the clinical-only, radiomics-only, composite, and EfficientNetV2 models for both endpoints (all P < 0.0001). It correctly identified patients likely to achieve PFS > 24 months and those unlikely to reach 12 months, which was superior to tumor mutation burden (TMB) and PD-L1 expression. Higher DHGN scores strongly predicted longer survival outcomes (hazard ratio: 0.10, 95% CI: 0.07-0.16, P < 0.0001) and improved the prognostic accuracy of patient stratification based on PD-L1 expression. Based on clinicians' practical deployment, the DHGN significantly reduced the operational complexity and computational requirements (P < 0.01) for model development and application in clinical settings. In conclusion, we proposed a clinician-friendly population graph model that fuses baseline clinical data with CT radiomics on widely available CPU hardware accurately stratifies NSCLC patients for immunotherapy benefits, potentially redefining benchmarks for non-invasive prognostic biomarkers and enabling broader clinical translation.