Maxime Mermod1, Eva-Francesca Jourdan2, Ruta Gupta3,4,5, Massimo Bongiovanni6, Genrich Tolstonog1, Christian Simon1, Jonathan Clark4,5,7, Yan Monnier8. 1. Department of Otolaryngology - Head and Neck Surgery, Head and Neck Tumor Laboratory, CHUV and University of Lausanne, Lausanne, Switzerland. 2. Consultant Statistician for the Head and Neck Tumor Laboratory, CHUV and University of Lausanne, Lausanne, Switzerland. 3. Department of Tissue Pathology and Diagnostic Oncology, Royal Prince Alfred Hospital, Sydney, New South Wales, Australia. 4. Sydney Head and Neck Cancer Institute, Chris O'Brien Lifehouse, Sydney, New South Wales, Australia. 5. Central Clinical School, University of Sydney, Sydney, New South Wales, Australia. 6. Department of Clinical Pathology, Institute of Pathology, CHUV and University of Lausanne, Lausanne, Switzerland. 7. South West Clinical School, University of New South Wales, Sydney, New South Wales, Australia. 8. Department of Otolaryngology - Head and Neck Surgery, Geneva University Hospital and Faculty of Medecine of the University of Geneva, Geneva, Switzerland.
Abstract
BACKGROUND: There have been few recent advances in the identification of occult lymph node metastases (OLNM) in oral squamous cell carcinoma (OSCC). This study aimed to develop, compare, and validate several machine learning models to predict OLNM in clinically N0 (cN0) OSCC. METHODS: The biomarkers CD31 and PROX1 were combined with relevant histological parameters and evaluated on a training cohort (n = 56) using four different state-of-the-art machine learning models. Next, the optimized models were tested on an external validation cohort (n = 112) of early-stage (T1-2 N0) OSCC. RESULTS: The random forest (RF) model gave the best overall performance (area under the curve = 0.89 [95% CI = 0.8, 0.98]) and accuracy (0.88 [95% CI = 0.8, 0.93]) while maintaining a negative predictive value >95%. CONCLUSIONS: We provide a new clinical decision algorithm incorporating risk stratification by an RF model that could significantly improve the management of patients with early-stage OSCC.
BACKGROUND: There have been few recent advances in the identification of occult lymph node metastases (OLNM) in oral squamous cell carcinoma (OSCC). This study aimed to develop, compare, and validate several machine learning models to predict OLNM in clinically N0 (cN0) OSCC. METHODS: The biomarkers CD31 and PROX1 were combined with relevant histological parameters and evaluated on a training cohort (n = 56) using four different state-of-the-art machine learning models. Next, the optimized models were tested on an external validation cohort (n = 112) of early-stage (T1-2 N0) OSCC. RESULTS: The random forest (RF) model gave the best overall performance (area under the curve = 0.89 [95% CI = 0.8, 0.98]) and accuracy (0.88 [95% CI = 0.8, 0.93]) while maintaining a negative predictive value >95%. CONCLUSIONS: We provide a new clinical decision algorithm incorporating risk stratification by an RF model that could significantly improve the management of patients with early-stage OSCC.