Replicating Patient Follow-Up with Hierarchical Directed Graphs for Head and Neck Cancer Survival Analysis

Théo Di Piazza, Hugo Miccinilli

Keywords: survival analysis, head and neck cancer, multimodal, graph neural networks

Abstract:

Head and neck cancer is a common malignancy with persistently limited survival outcomes, making accurate clinical prognosis particularly challenging. To establish a diagnosis, patients typically undergo a series of examinations producing heterogeneous data. This includes clinical data review, blood tests, tissue sampling, and lymph node analysis, encompassing multiple imaging and non-imaging modalities prior to treatment, which often involves surgery for disease treatment. Despite advances in diagnostic imaging and clinical assessment, treatment decisions remain largely dependent on the disease stage. This highlights the critical need for automated and reliable tools capable of accurately estimating patient survival to further assist clinicians in personalized treatment planning.
Existing survival analysis methods, typically rely on shallow architectures or early-fusion schemes that struggle to exploit the complexity and structure of multimodal clinical data. To address these limitations, we introduce H2DGSurv, a deep learning framework for survival prediction in head and neck cancer that models multimodal patient data as a directed hierarchical heterogeneous graph tracing the clinical workflow from initial diagnosis to surgery. The proposed architecture organizes modality-specific leaf nodes under clinical step-level parent nodes, and integrates a global patient node to capture consolidated representations prior to survival prediction.
Experimental results demonstrate that H2DGSurv substantially improves survival prediction performance compared with established baselines, while ablation studies confirm the importance of each model component.

Primary Subject Area: Detection and Diagnosis

Secondary Subject Area: Integration of Imaging and Clinical Data

Registration Requirement: Yes

Reproducibility: https://github.com/dpmc-lab/h2dg-surv

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 47