3336 - Development of Machine Learning Algorithms from Computed Tomography Images to Predict Catastrophic Airway Events in Patients with Cancers in the Aerodigestive Tract

H. Bacon¹, R. Daniel², M. McInnis³, C. McIntosh^1,4, C. J. Tsai^1,5, and C. M. Yao^2,6; ¹Department of Radiation Oncology, University of Toronto, Toronto, ON, Canada, ²Department of Otolaryngology-Head and Neck Surgery, University of Toronto, Toronto, ON, Canada, ³Joint Department of Medical Imaging, University of Toronto, Toronto, ON, Canada, ⁴Department of Medical Biophysics, University of Toronto, Toronto, ON, Canada, ⁵Department of Radiation Oncology, Princess Margaret Cancer Centre, University of Toronto, Toronto, ON, Canada, ⁶Princess Margaret Cancer Centre, University Health Network, Toronto, ON, Canada

Purpose/Objective(s): Patients with cancers in the aerodigestive tract including upper aerodigestive tract and intrathoracic tumors (both primary and metastatic) are at risk for developing severe and potentially life-threatening complications including hemorrhage or airway obstruction. Preliminary data indicate that up to 50% of patients die within 100 days following emergency airway procedures or emergent palliative radiotherapy. This study aims to develop and apply machine learning (ML) algorithms to detect patients at high risk of experiencing catastrophic bleeding or airway obstruction events, enabling pre-emptive early intervention. Materials/

Methods: The study has two parts: (1) building a machine learning model to identify patients at risk for catastrophic airway events (airway obstruction or hemorrhage); and (2) verifying the feasibility and utility of applying the algorithm prospectively. For the first part, we will include approximately 300 patients with upper aerodigestive tract tumors previously treated at our institution for airway obstruction or hemorrhage. We will establish a baseline time point and find the computed tomography (CT) scan closest to that date. An airway segmentation algorithm will be used on the CT images to identify the proximal airway and vessels. We will then score the overlap between tumor and airway/vessels using a literature standard. This will be used as input for an ML-based predictive model to identify patients at high risk for catastrophic events. We will evaluate the performance of the model before and after inclusion of clinical factors, and compare the model performance to gold standard radiologist scoring from existing literature. For the second part of the study, the algorithm will be prospectively applied to a cohort of patients with metastatic cancer involving the aerodigestive tract in a “silent mode” to verify the performance of the algorithm. Based on the retrospective validation and results of the “silent mode” implementation, we will work with clinicians and patient partners to develop a protocol for a prospective study.

Results: This is a trial in progress with no results available to date.

Conclusion: This study will aim to develop and prospectively validate an algorithm for automatic screening of patients at risk for developing catastrophic airway events. This may inform the application of prophylactic airway interventions to mitigate morbidity and mortality associated with hemorrhage or airway obstruction in patients with aerodigestive tract tumors.