Abstract
Background: Pre-treatment predictors of laryngeal preservation (LP) and survival in advanced laryngeal-hypopharyngeal squamous-cell carcinoma (LHSCC) represent an unmet clinical need. Materials and methods: A multicentric, international, retrospective series of LHSCC patients undergoing induction chemotherapy (IC) within an LP protocol was analyzed. The primary objective was to develop a predictive model by exploiting multiomics data (clinical, genomics, radiomics). Endpoints were laryngo-esophageal dysfunction (LED), response to IC, overall survival (OS), and progression-free survival (PFS). Patients were divided into three groups: group A (no LED); group B (responders to IC with LED); group C (non-responders to IC with LED). Several algorithms (support vector machine, random forest, C5.0, k-nearest neighbors, XGBoost, and naive Bayes) were run and compared in terms of multiclass area under the curve (AUC) score and classification error. Results: One hundred and ninety-one LHSCC patients were included (median age 60 years, 72% laryngeal, 80% T1-T3, and 58% N+). Responders to IC were 85%, while 66% suffered from LED. The 5-year PFS and OS were 58.4% and 64.7%, respectively. When comparing the three predictive models (clinical, clinical + genomics, clinical + radiomics), the addition of genomics provided the highest AUC. Then, we selected a 64-gene signature and 6 clinical variables (comorbidities, primary site, smoking, T category, N category, performance status) to build up the PRESERVE model. It showed a classification error of 28.9% and an AUC of 87.4%. Risks of major misclassification were low (group A to C, 1.13%; group C to A, 7.38%). Decision analysis confirmed the efficiency of the model. Conclusions: The PRESERVE model proved to be efficient and accurate in predicting LED and response to IC in LHSCC. External validation is needed before clinical application.