Can Omics Help in Prognostic Machine Learning Interpretability

Purpose/Objective(s) This study aims to address a question on the potential role of omics in machine learning (ML) interpretability. Materials/Methods We performed a retrospective study to investigate the association between clinical features and overall survival (OS) in 778 patients with stage III or IV head and neck squamous cell carcinoma (HNSCC), who participated in the RTOG 0522 randomized trial, by using an interpretable ML model known as the Explainable Boosting Machine (EBM). By extracting and adding omics (radiomics and dosiomics) features to the clinical features, we attempted to determine the potential role of omics with respect to the interpretability of the ML model for predicting OS. Results The EBM model showed clear details of the contribution of each feature to OS prediction and supported the primary findings in RTOG 0522 that the primary tumor site, N2b-3 category, T4 tumor, smoking history ≥ 10 pack-years of cigarette, age > 50 years, and Zubrod performance status are strong predictors of poor OS. The interpretable model also predicted a 2-year survival with 69% and a 5-year survival with 72% in accuracy in the test dataset, despite not having certain critical features such as human papillomavirus status due to significant missing values. The performance of the EBM model for predicting 2-year and 5-year survival achieved the area under the curve of 0.69 and 0.73, respectively. When dosiomics information was added to the dose-volume histogram parameters, we observed a clearer trend in the model's explanations on risk scores for clinical features, illustrating that omics could promote the development of personalized medicine. Conclusion Our findings contribute to the evidence that the addition of omics in prognostic models can improve ML interpretability.