Radiomics CoxPH Model for Overall Survival Prediction in Head and Neck Cancer

Head and neck cancer (HNC) includes a wide range of tumors that often behave very differently, even when they are classified under the same stage. In clinical practice, factors like TNM staging are commonly used to predict outcomes, but They do not always reflect the true complexity or biological behavior of the disease. Radiomics has emerged as a promising approach in this area. In this study, Computed Tomography (CT)-based radiomic features are used to develop a survival prediction model using the Cox Proportional Hazards (CoxPH) approach. The aim is to improve patient outcome prediction and to help with better risk stratification leading to better clinical decision making.

The study was approved as retrospective study by IEC of our institution. In total 107 patient were included in this study who were diagnosed with squamous cell carcinoma in the head and neck region. In total 103 IBSI-compliant original radiomic features were extracted from the primary tumor using the PyRadiomics library. Multistep feature selection methods were used. As first step correlating features were removed based on the Pearson correlation method (correlation coefficient >0.7) and as a second step multivariate Cox Proportional Hazards (CoxPH) model was used to select optimal features. Subsequently the survival analysis was performed using a CoxPH model implemented through the lifelines Python package. Model performance was assessed using the Concordance Index (C-index), while Kaplan–Meier curves were used to visualize survival differences between risk groups.

The final CoxPH model included 13 radiomic features spanning shape, first-order, and texture categories. The model showed strong performance: Training Set: C-index = 0.68, Independent Test Set: C-index = 0.86. Using the risk scores predicted by the CoxPH model, patients were divided into high-risk and low-risk groups based on the median score.

This study successfully developed an IBSI-compliant radiomics pipeline combined with a Cox Proportional Hazards model predicts survival in head and neck cancer patients. The CoxPH model demonstrated strong discriminative performance in separating patients into clinically meaningful risk groups.