HNC-PREDICTOR

HNC-PREDICTOR predicts patient-specific risk for head and neck cancer (HNC) to support personalized radiotherapy planning through multivariate prognostic modeling.

Key Features:

• Data-Driven Model Development: Trained on data from 4,611 HNC patients collected across three academic cancer centers using a multi-cohort strategy with cohorts split into training (n = 2,241), independent test (n = 786), and external validation (n = 1,087 and n = 497).
• Machine Learning Pipeline: Employs a machine learning pipeline that integrates tumor- and patient-related clinical variables to predict overall survival as the primary endpoint and local and regional tumor control as secondary endpoints.
• Predictive Variables: Identified key predictors include performance score, AJCC 8th stage classification, pack-years of smoking history, and age, with demonstrated predictive performance across validation cohorts.
• Risk Stratification: Stratifies patients into high-, intermediate-, and low-risk groups to inform radiotherapy decisions such as dose escalation or de-escalation.
• Imaging Feature Integration: Inclusion of morphological image features improved predictive accuracy, achieving a c-index of 0.73 [95% CI, 0.64–0.81].

Scientific Applications:

• Personalized Radiotherapy Planning: Enables tailoring of radiotherapy regimens (e.g., dose escalation or de-escalation) based on individualized risk profiles.
• Prognostic Stratification: Provides prognostic subgrouping that distinguishes high-risk patients (reported 5-year overall survival 17–46%) from low-risk patients (reported 5-year overall survival 92–98%).
• Clinical Decision Support for Treatment Selection: Supports selection of patients for intensified or de-intensified treatment strategies and for enrollment in risk-stratified clinical studies.

Methodology:

Models were developed using a multi-cohort machine learning approach trained and validated on a dataset of 4,611 patients (training n = 2,241; independent test n = 786; external validation n = 1,087 and n = 497), integrating tumor- and patient-related clinical variables to predict overall survival (primary) and local/regional control (secondary), with optional incorporation of morphological imaging features.