AI voice analysis may help doctors spot early signs of ALS, study suggests

Approach may be particularly useful in form that initially affects mouth and throat

Written by Marisa Horak, MS |

The screens of a computer and laptop show 0s and 1s.
  • AI voice analysis shows potential in detecting early signs of ALS, particularly the bulbar-onset form affecting the mouth and throat.
  • Bulbar-onset ALS carries a poorer prognosis, making early diagnosis particularly important.
  • Machine learning models, such as XGBoost, distinguished ALS patients from healthy individuals with high accuracy.

Analyses of vocal recordings using artificial intelligence (AI) may aid in the diagnosis of amyotrophic lateral sclerosis (ALS), particularly among those with a form of the disease called bulbar-onset ALS that initially affects the mouth and throat, a new study suggests.

Bulbar-onset ALS generally has a poorer prognosis than limb-onset disease — in which symptoms of muscle weakness initially affect the arms and legs — making early diagnosis particularly important. In this study, a machine learning model distinguished people with ALS from healthy individuals with 96% accuracy, although the findings came from a small data set and need to be validated in larger groups.

“This paper highlights the potential of AI-based approaches in assisting medical researchers and clinicians in the early detection and treatment of Bulbar ALS,” researchers wrote in the study, “Meta-learning approach for ALS detection using sustained vowel phonations,” which was published in Discover Artificial Intelligence.

Recommended Reading
A clinician holding a clipboard gestures while speaking to a patient sitting on an examining table.

Blood test may improve accuracy and reduce false ALS diagnoses

Bulbar-onset ALS can cause speaking difficulty

ALS is a neurological disease that causes progressive muscle weakness. In some people with ALS, the disease initially affects muscles in the throat and around the mouth. This is known as bulbar-onset ALS, and it can cause symptoms such as difficulty eating and speaking.

Machine learning is a form of AI that basically involves feeding a computer a large set of data, alongside mathematical rules that the computer uses to identify patterns within the data. These patterns can then be used to make sense of future data sets.

Because bulbar-onset ALS often changes how people speak early in the disease course, researchers have posited that it might be possible to detect the disease by using machine learning to analyze vocal recordings. In this study, a pair of scientists in India put that idea to the test.

By leveraging advanced techniques and analyses of audio data, this research contributes to the ongoing efforts in improving the understanding and management of this condition.

The researchers used an existing data set including vocal recordings from 64 people, roughly half of whom had ALS. Among patients, 13 had bulbar-onset ALS and 18 had limb-onset disease, although some of the latter also exhibited signs of bulbar symptoms.

For each participant, two vocal recordings were available: one of them saying the sound /a/ (the vowel sound in cat or bat), and another of them saying the sound /i/ (the sound in pig or big). The researchers extracted specific sound-based measures from these recordings, then evaluated dozens of different machine learning paradigms to look for ones that could accurately distinguish the people with ALS from healthy controls.

“By leveraging advanced techniques and analyses of audio data, this research contributes to the ongoing efforts in improving the understanding and management of this condition,” the scientists wrote.

Recommended Reading
In this rare disease illustration, a single red balloon is seen among more than a dozen black and white balloons.

Researcher honored for AI brain implant that helps restore speech in ALS

One machine learning model reached 96% accuracy

The team evaluated 24 machine learning algorithms and a deep-learning model to determine how accurately they could distinguish people with ALS from healthy controls. One model called XGBoost performed best among the machine learning approaches, reaching 96% accuracy.

The researchers also tested few-shot learning approaches, which are designed to work with relatively small amounts of labeled data. Among these approaches, a prototypical network performed best, reaching nearly 87% accuracy.

The findings suggest that acoustic features from sustained vowel sounds may contain patterns that can help distinguish people with ALS from healthy individuals. However, the results are preliminary and additional studies will be needed to validate and expand on these results.

Still, it’s possible that voice-based analyses using cutting-edge AI may one day help clinicians to diagnose people with ALS. This type of approach “offers a non-invasive and potentially efficient method for identifying the disease at an earlier stage,” the team wrote.

Leave a comment

Fill in the required fields to post. Your email address will not be published.

Comments are moderated. Once approved, your comment and username will be publicly visible. Please avoid sharing personal health information or other sensitive details.