Abstract
Traditional survival predictions for oral squamous cell carcinoma (OSCC) rely on TNM staging, which lacks individualized prognostic value. Clinical factors such as performance status, age, sex, and lifestyle affect outcomes but are underrepresented in conventional models. This study applied artificial intelligence (AI) to integrate diverse factors for OSCC survival prediction. We retrospectively analyzed 1,018 OSCC patients surgically treated between 1996 and 2020. Variables included demographics, lifestyle, ASA classification, TNM stage, PET SUVmax, peri-neural and lympho-vascular invasion, extranodal extension, depth of invasion, resection margin, and treatment modalities. A deep neural network (DNN) for multi-group classification was developed and compared with regression-based DNN, Cox proportional hazards, and random survival forest models. To address class imbalance, least squares and multi-task learning were applied. Performance was assessed with concordance index and linearity testing. Death occurred in 18.1% of patients, with mean survival of 36.8 months. Recurrence occurred at 33 months. The DNN achieved an AUC of 0.922, sensitivity 0.514, specificity 0.992, and concordance index 0.888. Linearity testing confirmed strong correlation between predicted and observed outcomes. AI models integrating clinical variables provide more accurate OSCC survival predictions than conventional staging. The multi-group DNN is a promising tool for individualized prognosis and treatment planning.