Machine learning explainability for survival outcome in head and neck squamous cell carcinoma

Area of Science:

• Oncology
• Machine Learning in Medicine
• Cancer Prognostics

Background:

• Head and neck squamous cell carcinoma (HNSCC) diagnosis and treatment significantly impact patients psychologically and cause treatment toxicity.
• Evaluating patient outcomes is crucial for effective treatment planning and improved disease management in HNSCC.

Purpose of the Study:

• To develop a prognostic machine learning (ML) model integrating clinicopathological, treatment-related, and sociodemographic factors to predict overall survival (OS) in HNSCC patients.
• To explore the complementary prognostic potential of these diverse input parameters.
• To provide model explainability using Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP).

Main Methods:

• Recruited 419 HNSCC patients from three Swedish University Hospitals.
• Compared the performance of TabNet, extreme gradient boosting (XGBoost), and a voting ensemble for OS prediction.
• Utilized LIME and SHAP for model interpretability.

Main Results:

• TabNet and XGBoost achieved comparable prediction accuracies of 88.1%, with the voting ensemble reaching 88.7%.
• Key predictors for OS included p16 status, cancer stage, hemoglobin, age, T and N class, smoking history, BMI, treatment, erythrocyte count, and HPV status.
• Analysis revealed survival trends associated with p16, cancer stage, hemoglobin, age, HPV status, TNM staging, and socioeconomic factors.

Conclusions:

• Clinical implementation of ML models can facilitate individualized, risk-based therapeutic decision-making for HNSCC patients.
• Further validation using multi-institutional datasets and clinical trials is necessary for safe clinical implementation of these prognostic models.