A Pilot Study to Characterize the Voice Signatures Associated with Early Glottic Cancer

Dr. Baird hopes to build a high-throughput screening tool utilizing a large amount of audio data already collected to diagnose, stratify and triage people at risk for laryngeal cancer. Laryngeal cancer is the second most common head and neck cancer in the US and catching cancer progression early is critical to both survival and preserving the quality of a patient’s life, but currently the best way to catch it is by referral to a specialist, often outside of regular insurance coverage. Head and neck oncologists need a better and less expensive way to monitor those at risk. Thus far, AI has been successfully utilized to interpret imagining data to identify stages and risk of cancer; to this Dr. Baird hopes to add multi-heuristic voice recordings. He has preliminary data from an already existing collaboration with the POLIMI Research Institute in Italy that suggests that the model he is building will be able to accurately distinguish between sounds made by a person with cancer or dysplasia and one who has normal vocal folds.

Up until now, attempts to identify audio clues to cancer development have focused on specific distinct units of sound. Dr. Baird has confidence that by using a much larger whole sound data set paired with better noise filtering and processing, he will be able to train a tool that will be able to perform reliable voice-based diagnostics. In his test data set, early versions were able to distinguish sounds from patients with cancer from those with normal vocal folds almost 80% of the time. Using this Young Investigator Award, Dr. Baird will continue to refine his model, hoping to bring his sound identification tool to a point where it can not only distinguish between the sounds made by cancerous and normal vocal cord tissue, but also identify other pathologies like polyps or vocal fold paralysis. Ideally this model may also provide a way to scan for cancer post-surgery, providing a huge benefit to patients monitoring for reoccurrence.