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

Abstract

Background. Diagnosis and treatment of head and neck squamous cell carcinoma (HNSCC) induces psychological variables and treatment-related toxicity in patients. The evaluation of outcomes is warranted for effective treatment planning and improved disease management. Objectives: This study aimed to build a prognostic system by combining clinicopathological parameters, treatment-related factors, and sociodemographic factors as integrative inputs to build a machine learning (ML) model to estimate the overall survival (OS) of patients with HNSCC. Furthermore, we explored the complementary prognostic potentials of these input parameters. We provide explainability and interpretability using Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) techniques. Methods: A total of 419 patients with HNSCC were recruited from three University Hospitals in Sweden. We compared the performance of TabNet, a state-of-the-art deep learning algorithm for tabular data, with extreme gradient boosting (XGBoost) and voting ensemble to predict OS in patients with HNSCC. Results: Both TabNet and XGBoost showed comparable performance accuracies, with TabNet and XGBoost showing a performance accuracy of 88.1% each and voting ensemble showing an accuracy of 88.7%. The aggregate feature importance showed that p16 (a tumor suppressor protein that plays a crucial role in cell cycle regulation), cancer stage, hemoglobin, age at diagnosis, T class, N class, smoking pack-years, body mass index (BMI), treatment modality, erythrocyte count, and human papillomavirus (HPV) status were the most important parameters for the predictive ability of the model for OS. Furthermore, we found survival trends in this cohort by individually considering parameters such as p16, cancer stage, hemoglobin, age at diagnosis, HPV status, Tumor Nodal Metastasis staging, and socioeconomic factors (marital status, housing, and level of education). In addition, both the LIME and SHAP techniques showed the contribution of each feature to the prediction made by the model. Conclusions: The clinical implementation of an ML model can lead to individualized risk-based therapeutic decision-making. Therefore, validating these models with multi-institutional datasets and testing them in the context of clinical trials is warranted for safe clinical implementation.