2026 Poster Presentations
P310: DEVELOPMENT OF A PRECISION MEDICINE MODEL FOR SINONASAL CANCER USING ARTIFICIAL INTELLIGENCE
Edmund Zhi, BS1; Bishoy M Galoaa, MS2; Andrew G Girgis, BS1; Andrew Tsao, BS1; Anand K Devaiah, MD3; 1Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States; 2Northeastern University College of Engineering, Boston, MA, USA; 3Department of Otolaryngology/Head and Neck Surgery, Boston Medical Center, Boston University School of Medicine, Boston, MA
Importance/Background: Integration of artificial intelligence (AI) in healthcare has been rapidly expanding. A large opportunity for this technology is using AI for precision medicine in rare and complex diseases. Sinonasal cancers are a broad, heterogenous group of malignancies which frequently present significant challenges for clinicians to diagnose and manage due to their rarity and diversity, variability around treatment regimens, and challenges in conducting large-scale studies. Thus, we aimed to develop an AI model and explore its potentials in assisting clinicians in personalized decision-making in sinonasal cancer management.
Objective: Our model aims to identify the most important prognostic factors to predict and tailor optimal treatment regimens to maximize survival for sinonasal cancer patients to create a precision medicine approach and further personalize treatment.
Design, Setting, and Participants: Our cohort included 3,859 patients diagnosed with sinonasal cancer from the Surveillance, Epidemiology and End Results (SEER) dataset. 33 features comprising of demographic information, clinical characteristics, treatment modalities, and pathological findings from electronic medical records were extracted, processed, and standardized. Feature importance was automatically calculated and utilized in a robust neural network architecture which included a multi-layer perceptron with dropout layers (p=0.2) for regularization, batch normalization between fully connected layers, ReLU activation functions in hidden layers and sigmoid activation in the output layer for probability estimation. Model training included a binary cross-entropy loss function and early stopping based on validation loss. The dataset was randomly split into training (80%) and testing (20%) sets. Model performance was measured through AUC-ROC, sensitivity, specificity, and accuracy. Statistical analysis was further performed through 5-fold cross-validation, DeLong's test for comparing ROC curves, and Kaplan-Meier analysis for survival outcomes.
Results: We evaluated two hybrid survival models, with and without demographic factors. Our model without demographic features produced a time-dependent AUC of 0.799 at 12 months, 0.820 at 24 months, 0.922 at 60 months, and 0.829 at 120 months. Our model with demographic features produced a time-dependent AUC of 0.808 at 12 months, 0.832 at 24 months, 0.921 at 60 months (highest discriminating ability), and 0.846 at 120 months. The top 10 prognostic features included established features and others related to Social Determinants of Health (SDoH) and are (descending order): tumor size, metastases at diagnosis, median household income, metastases to distant lymph nodes, age, surgery performed on primary site, time from diagnosis to treatment, year of diagnosis, histology characteristics, and T stage.
Conclusions: Our results indicate that the developed model has significant potential for demonstrating clinical relevance and assisting clinicians in the management of sinonasal cancers. This model is able to identify the most important prognostic features and then use these to provide a precision medicine personalized regimen to maximize individual patient survival.