Interpretable survival network for progression risk analysis of multimodality imaging biomarkers in poor-prognosis...

This study explores the predictive utility of multi-time point, multi-modality quantitative imaging biomarkers (QIBs) and clinical factors in patients with poor-prognosis head and neck cancers (HNCs) using interpretable machine learning. We examined 93 patients with p16 + oropharyngeal squamous cell carcinoma or locally advanced p16- HNCs enrolled in a phase II adaptive radiation dose escalation trial. FDG-PET and multiparametric MRI scans were conducted before radiation therapy and at the 10th fraction (2 weeks). A survival network analyzed MRI and PET-derived biomarkers such as gross tumor volume (GTV), blood volume (BV), and metabolic tumor volume (MTV₅₀), along with clinical factors to predict local (LF) and distant failures (DF). Feature attributions and interactions were assessed using Expected Gradients (EG) and Expected Hessian (EH). Through rigorous cross-validation, the model for predicting LF, incorporating biomarkers like p16 status and radiation boost, achieved a c-index of 0.758. Similarly, the DF prediction model showed a c-index of 0.695. The analysis of feature attributions and interactions enhanced understanding of important features and complex factor interplays, potentially guiding more personalized and intensified treatment approaches for HNC patients.