Voice-Based Screening Predicts Lung Function

Results from a preliminary study showed novel voice and breath analysis using a smartphone app to be a useful measure of lung function, and researchers say the technique may prove to be a valuable tool for identifying and monitoring respiratory disease.

Automated voice and breath analysis was found to have value for predicting lung function, with an 82% accuracy for predicting patients with and without obstructive lung disease shown in the study, which was presented at the virtual CHEST conference, the annual meeting of the American College of Chest Physicians.

Since breath is important for controlling the dynamics of speech, we hypothesized that voice could be a useful marker of lung function, said lead author Obaid Ashraf, MD, of Allegheny Health Network in Pennsylvania, who presented the findings.

While spirometry is the gold standard for measuring lung function, there are significant drawbacks to using it in clinical practice, he said, including the need for specialized equipment that must be maintained and sterilized and the need for skilled technicians to administer it.

In addition, spirometry is not readily available in the primary care setting, and it is often difficult to obtain reproducible spirometry results from patients with lung diseases and comorbid conditions.

"Spirometry is a great tool. I love using it. I want to get it in every patient I see in house and in the clinic, but that is just not realistically possible," Ashraf said. "The question was, can we use something else that is more convenient to give us some idea what is happening?"

The ongoing, prospective, cross-sectional study included 128 initial participants (76 women and 52 men), who were recruited during appointments for regularly scheduled pulmonary function testing conducted at Allegheny General Hospital in Pittsburgh. Of the cohort, 16.4% had lung obstruction.

A voice collector app, designed by the researchers, was used to collect voice data. Participants were asked to read a phonetically-balanced passage with 199 words selected to optimize findings. The sample passage was later reduced to around 50 words.

Voice and breath sound samples were recorded before and after pulmonary function testing, which corresponded with pre-and post-bronchodilator samples.

The researchers obtained pre- and post-pulmonary function test forced expiratory volume in one second (FEV1) and forced vital capacity (FVC) results from each patient. Participant voice and breath audio samples were recorded on a smart tablet, using the proprietary software app. The recorded voice data were analyzed using cloud-based software. Voice audio recordings were calibrated to create a customized noise profile.

The phonetically balanced reading passage was used to examine respiration, phonation, articulation, and resonance, while a long vowel word list was used to detect speaking-related dyspnea during articulation of long vowel sounds.

Machine learning was used to compare the voice-based screening to spirometry data.

In his presentation, Ashraf reported that the automated voice analysis delivered good diagnostic accuracy for the prediction of FEV1 (R squared 0.74, mean squared error 0.17, mean absolute error 0.32, binarized accuracy 71.5%) and FVC (R squared 0.79, mean squared error 0.19, mean absolute error 0.35 binarized accuracy 71.4%).

Obstruction classification showed an accuracy of 98%, a sensitivity of 96%, and an F1 score of 0.9, he said.

The researchers plan to conduct a longitudinal study in patients with chronic obstructive pulmonary disease as a next step in their research to determine if voice monitoring can predict exacerbation risk.

Ashraf said the technique "offers the promise of widespread, affordable medical screening and real-time monitoring of respiratory disease," though he acknowledged during his presentation that the clinical value of voice-based screening in the diagnosis and management of patients with chronic respiratory disease is not yet known.

In the absence of spirometry, "some information (on lung function) is better than no information," he noted.

"We started with that idea, and I think that this is a work in progress that we can build on," he said.