What's in a voice?

A doctor works a lot on intuition. A patient walks in, seemingly fine, observations all normal but there’s something not quite right. They’re just a little bit vague, a little too slow, a bit… off. For some reason you pursue the investigations, struggling to justify why exactly you want it done. Finally, the patient opens up about the odd things they’ve been experiencing: voices, images, thoughts that confuse and frighten them. You suspect a serious mental illness, you refer to psychiatry.

It’s a struggle to teach the spidey sense in medical school. Given that doctors sometimes miss even a glaringly obvious diagnosis, trying to teach intuition seems a bit redundant. But what if there were technologies that could codify these subtle signs and help make a diagnosis?

This is what the emerging field of vocal biomarker technology is trying to do. Particularly useful in mental health disorders, these software programs analyse speech, looking for subtle signs of disease in the things you say and the way you say it. These software programs analyse speech, looking for subtle signs of disease in the things you say and the way you say it.

“These software programs analyse speech, looking for subtle signs of disease in the things you say and the way you say it.”

Forging ahead is Psyrin, a startup using AI to analyse something as short as 5 minutes of speech to help diagnose serious mental illnesses (SMIs) like schizophrenia. Considering a lot of the bottleneck in initiating treatment is the diagnosis of a condition (bipolar, for example, takes 9.5 years to diagnose according to bipolar UK), improving diagnostic technologies could be the key to shaving down this time. But will it actually work? How does this fit into the system as it stands?

What can vocal biomarkers add to the landscape

Mental health has always been a bit nebulous in its diagnosis. Unlike the obvious signs of physical disease e.g. a blocked artery or a TB bacillus under the microscope, diagnosing mental health conditions is less clear cut. Of help are checklists and questionnaires e.g. the PHQ-9 for depression, however, these are not prerequisite. A lot of the time, a diagnosis lies in a doctor’s opinion.

Enter the vocal biomarker. If the evidence backs it up and the technology can reliably support a diagnosis of a mental health condition, vocal biomarkers could change the SMI pathway. It could finally mean a quantifiable marker for mental health.

The technology is already being developed for neurological disorders like Parkinsons and myasthenia gravis with companies like BioSensics Biodigit and Canary Speech developing AI-assisted tools that can monitor and predict flares in a wide range of conditions based.

If vocal biomarker techonology works, it could be a marvel in terms of ease of measurement. Imagine this - a voice-recorded questionnaire done by the patient in their own home that gives an accurate reading of their mental health. But is it too good to be true?

Does it work?

I’m always a bit wary of hearing “AI-magic” as a term to describe how a medical software comes to important conclusions. I want to know what processes went into turning 5 minutes of speech into a possible diagnosis of schizophrenia. From what I understand, by training an AI on a previous validated dataset (in Psyrin’s case 20 000 minutes of voice recordings from patients with various mental health conditions), the trained AI can then predict outcomes from new data (e.g. the 5 minutes of speech) it encounters.

The proof of the pudding will be in having clinical trial outcomes. Sonde Health, a US based company using vocal biomarkers, has shown promising results with a recent study published in Frontiers in Psychiatry. The 104 participant study showed there was a strong correlation between their software’s mental health scoring system and the standard mental health assessment questionnaire used to judge symptoms. Importantly, there was suggestion that heightened awareness of the patient’s mood scoring, allowed them to change their behaviour to improve their mental health.

Obviously more works needs to be done, but there is a promise.

Pitfalls

It’s easy for me to poke holes, but I think it’s important to discuss the limitations of any new med tech device. I’m always sensitive to variations in patient populations and wonder how well the software can account for differences in speech patterns, dialect and language. Obviously, if that introduces inaccuracies into the diagnosis, then that’s a major flaw in the product. Furthermore, if the software can only be used for people who fit their training dataset, this could further heighten health inequality and access to services, something that in these stretched times we need to be even more mindful of.

In summary, the potential for vocal biomarker technology in the diagnosis and even the management of mental health conditions could prove revolutionary. However, as with all med tech products, it needs to be validated, safe and inclusive to really break new ground.