What your voice says about your smoking habits

LIH identifies a vocal biomarker of smoking status

• Deep Digital Phenotyping Research Unit
• Colive Voice

Building on their flagship Colive Voice research programme, researchers from the LIH Deep Digital Phenotyping Lab at the Department of Precision Health (DoPH) have developed a novel digital vocal biomarker that can be used in clinical and epidemiological research to assess smoking status in a rapid, scalable and accurate manner using audio recordings. This innovative technology displays the ability to differentiate smokers from non-smokers with good accuracy, regardless of their gender and spoken language.

Smoking is a major global public health issue and the leading preventable cause of illness and death, being responsible for over eight million deaths annually, including 1.3 million non-smokers affected by second-hand smoke. The negative health effects of smoking include immediate issues like oxidative stress and long-term consequences such as a 2- to 4-fold increase in the risk of coronary heart disease and stroke, the development of at least 12 types of cancer, and ultimately, the death of half of those who do not quit. For these reasons, smoking status is one of the main variables evaluated in clinical or epidemiological research, for instance when monitoring treatment response. However, frequent inconsistencies, data gaps and biased information call for additional methods to objectively assess smoking status.

“Voice features hold great potential as a tool to assess smoking status. Indeed, smokers are more likely to experience voice changes than non-smokers, as smoking affects various aspects of vocal function and quality”, says Hanin Ayadi from the Deep Digital Phenotyping lab at DoPH. “We therefore aimed to analyse the effects of tobacco smoke on voice and develop a gender and language-specific vocal biomarker for the automatic identification of the smoking status”, she explains.

The research team leveraged advanced data analysis and AI methods to analyse 3,996 voice recordings from 1,332 participants from the flagship Colive Voice research programme. Interestingly, they showed that smoking has more significant effects on the female voice compared to the male voice. Indeed, for women, smoking lowered the voice’s fundamental frequency, i.e. the number of times per second that vocal cords vibrate when making voiced sounds, as well as the mean harmonic-to-noise ratio, a measure that quantifies the amount of noise in the voice signal, in addition to other features associated with different vowel quality in speech sounds. This indicates that smoking affects the resonance of the vocal tract and proper articulation. Conversely, fewer differences between smokers and non-smokers were noted in men. The greater vulnerability of the female voice to smoke could be explained by anatomical and physiological differences between the two genders including, for women, shorter vocal cords and greater hormonal fluctuations that could affect the larynx, among others.

Subsequently, by combining a series of voice features, the researchers successfully managed to predict smoking status with an accuracy of up to 72% for English-speaking women and up to 72% for French-speaking men, thus providing a gender and language-specific vocal biomarker to distinguish between smokers and never-smokers. “In essence, we demonstrated that the effect of smoking on the voice displays gender- and language-specific disparities, which is what we took into account when devising different combinations of vocal  features  to  develop our novel vocal biomarker”, summarises Dr Guy Fagherazzi, Director of the DoPH and corresponding author of the study.

Our work lays the foundations for the automatic identification of smoking habits from voice recorded in real-life settings, which provides an efficient, non-invasive and inexpensive way to monitor this unhealthy practice in clinical and epidemiological research studies,

he concludes.

The study was published in the journal “Digital Biomarkers” with the full title “Digital vocal biomarker of smoking status using ecological audio recordings: results from the Colive Voice study”.