Explainable Boosting Machine Model with a Parallel Ensemble Design Predicts Local Failure for Head and Neck Cancer With Clinical, CT, and Delta CBCT Radiomic Features.

PURPOSE/OBJECTIVE(S) Local failure (LF) following chemoradiation for head and neck squamous cell carcinoma (HNSCC) is associated with poor overall survival. If machine learning techniques could stratify patients at risk of treatment failure based on baseline and intra-treatment imaging, such a model could facilitate response-adapted approaches to escalate, de-escalate, or switch therapy. Our hypothesis was that adding baseline CT (CT1) and delta CBCT radiomic features to a clinical model would improve its discriminatory ability at predicting LF. MATERIALS/METHODS Participants treated at a single institution with definitive radiotherapy (RT) for HNSCC and who failed locally in-field at a primary or nodal structure (analyzed separately) were included and matched to controls (30 cases, 60 controls) based on the site of primary, T/N stage, and p16 status. Radiomic features were extracted from CT1 and CBCT scans at fractions 1 (CBCT01) and 21 (CBCT21; where CBCT21 - CBCT01 = delta features) of RT with open-source software and were selected for by: reproducibility (intra-class correlation coefficients < 0.95), redundancy (maximal relevance and minimal redundancy), and informativeness (recursive feature elimination). Separate models predicting LF of primaries or nodes were created using the explainable boosting machine classifier with 5-fold cross-validation for (1) clinical only, (2) radiomic only (CT1 and delta features), and (3) fused models (clinical and radiomic models fused). Twenty-five iterations were performed and predicted scores were averaged with a parallel ensemble design. Receiver operating characteristic (ROC) curves were compared against each other with paired-samples T tests in a statistical software with the ROC analysis function. RESULTS The fused ensemble model for primaries achieved an area under the curve (AUC) of 0.871 with a sensitivity of 78.3% and specificity of 90.9% at the maximum Youden J statistic, which trended towards improvement when compared to the clinical (AUC = 0.788, P = 0.134) but only reached significance when compared to the radiomic ensemble model (AUC = 0.770, P = 0.017). For nodes, the fused ensemble model achieved an AUC of 0.910 with a sensitivity of 100.0% and specificity of 68.0%, which also trended towards improvement when compared to the clinical model (AUC = 0.865, P = 0.080) but did not reach significance. The most commonly included radiomic features in the ensemble were delta shape features, with change in maximum 3D diameter and change in sphericity being the most common for the primary (69.6%) and nodal (80.8%) models. CONCLUSION The fused ensemble explainable boosting machine model achieved high discriminatory ability at predicting LF of HNSCC in independent primary and nodal structures. Although an additive benefit of delta radiomics over clinical factors could not be proven, the results trended towards improvement with the fused ensemble model, which are promising and worthy of further investigation in a larger cohort.