AI 'could detect laryngeal cancer from patient voices'

Researchers in the US have found that subtle changes in the sound of a voice could be used to detect if someone has cancer of the larynx, and even distinguish cancer from other disorders affecting the voice box.

The team, led by Dr Phillip Jenkins, a postdoctoral fellow in clinical informatics at Oregon Health & Science University, have said that their findings suggest AI could be used to analyse voice patterns as an alternative to current invasive diagnostic procedures for laryngeal cancer, like video nasal endoscopy and biopsies.

There is a wrinkle to the research in that the analysis was able to detect cancerous changes in men only. However, the scientists think that detecting the cancer in women could be possible if a larger dataset is used to train the AI. They have published their research in the journal Frontiers in Digital Health.

Jenkins said the results are a 'proof-of-principle' that vocal biomarkers could be used to identify if patients have lesions on the vocal fold of the larynx – also known as the vocal cords.

He and his colleagues drew on part of the NIH's Bridge2AI database, part of a national drive in the US to find biomedical applications for AI, which includes more than 12,500 voice recordings from around 300 participants.

The study focused on differences in a number of acoustic features of the voice, and found that the harmonic-to-noise ratio (HNR), a measure of the relation between harmonic and noise components of speech, and fundamental frequency or pitch were the most useful in distinguishing men without any voice disorder, men with benign vocal fold lesions, and men with laryngeal cancer.

The next step is to train these algorithms on more data and test them in clinical settings on patient voices, according to the team.

"Our results suggest that ethically sourced, large, multi-institutional datasets like Bridge2AI‑Voice could soon help make our voice a practical biomarker for cancer risk in clinical care," said Jenkins.

"To move from this study to an AI tool that recognises vocal fold lesions, we would train models using an even larger dataset of voice recordings, labelled by professionals," he added. "We then need to test the system to make sure it works equally well for women and men."

Various studies have looked at the potential of AI to detect other diseases from vocal biomarkers, including chronic obstructive pulmonary disease (COPD), dementia, amyotrophic lateral sclerosis (ALS), post-traumatic stress disorder (PTSD), depression, and stress.

Photo by Jelena Kostic on Unsplash