Enlarge image

Important biomarker:

The human voice as the “new blood” – or as the new gold mine for many start-ups

Photo: Tirachard / Getty Images

“The voice is the new blood,” is the headline of the German AI start-up Audeering in a blog entry: Be it Alzheimer's, depression or Parkinson's disease - the detection of diseases using AI voice analysis is now being promoted by many start-ups worldwide, including big names like IBM is getting involved. Computer scientist

Holger Fröhlich

(48) is familiar with the fusion of AI and medicine.

from: He heads the “AI and Data Science” department at the Fraunhofer Institute for Algorithms and Scientific Computing.

His focus is, among other things, on the development of AI methods that are used in medicine. Fröhlich is currently leading a federally funded project in which he is investigating whether AI can recognize the symptoms of Parkinson's just as well as a neurologist - and whether the technology can help to correctly record the course of the disease.

manager magazin: Mr. Fröhlich, various start-ups promise the early detection of diseases using AI voice analysis. How do you rate that?

Holger Fröhlich:

What you can recognize with such AI tools are the symptoms of a possible illness, not the illness itself. There are a variety of illnesses that can lead to language changes. Nevertheless, appropriate AI tools have their value when you look at the challenges facing our healthcare system. Not everyone can constantly run to a specialist and have incredibly expensive diagnostics carried out.

“What you can detect with such AI tools are the symptoms of a possible disease, not the disease itself.”

Why is voice the focus of companies?

It is known that voice changes are often associated with certain neurological and psychiatric diseases. In the project I lead, we are specifically investigating language changes in Parkinson's disease. But there are also other so-called digital biomarkers than voice changes: We also research gait and facial expressions.

How exactly does voice analysis using AI work?

One option is to use large transformer models that can recognize patterns in large amounts of data. If the aim is to distinguish sick from healthy, the AI ​​is fed with voice recordings of sick and healthy people and trained to distinguish certain patterns in the voice. For example, quiet and slurred speech is characteristic of Parkinson's disease. But what ultimately determines whether the AI ​​assesses someone as sick or healthy can only be understood to a limited extent. The AI ​​remains

- as it has often been described -

a bit of a black box.

How do you assess the situation on the market right now?

I think that's still one at the moment

exploratory initial phase in which we see many interesting approaches. Only time will tell what will ultimately become established in medical practice.

Enlarge image

AI expert:

Holger Fröhlich is group leader AI & Data Science at the Fraunhofer Institute for Algorithms and Scientific Computing

Photo: private

IBM and Pfizer have developed an AI model that is intended to help

detect possible Alzheimer's disease, and the US start-up Canary Speech wants to detect Alzheimer's disease using AI voice analysis. How promising is that?

Alzheimer's, as a biologically defined disease, cannot be recognized using language alone. This is not possible according to the clinical definition. What you can recognize through language are the symptoms of an incipient illness, which could possibly be Alzheimer's.

What are these symptoms?

The language becomes slower, at a certain point word-finding problems begin, the sentence constructions are somewhat strange, they do not correspond to the usual grammar. AI models could be used to capture symptoms that occur at an early stage of the disease. For many neurological diseases, the possible success of a treatment depends critically on when it is started.

The German AI company Audeering promises, among other things, the detection of Parkinson's symptoms using AI voice analysis.

The same applies here: Only a neurologist can make a definitive diagnosis because the disease is complex and there are a number of diseases that initially lead to similar symptoms. To my knowledge, there are currently no AI-based tools that can reliably distinguish language changes from all conceivable underlying diseases.

“To my knowledge, there are currently no AI-based tools that can reliably distinguish language changes from all conceivable underlying diseases.”

For example, depression can also be detected using AI voice analysis and emotion recognition. What do you think?

It is generally known that voice changes are a symptom of depressive illnesses. In this respect, I believe that a machine can be trained to detect these symptoms. I would be careful with emotional AI – depression is not an emotional state, but an illness.

In principle, in your opinion, it would make sense to use AI voice analysis for the early detection of some diseases. What might that look like in practice?

You have to imagine this as a process: Patients could use such AI tools at home, and if there are any abnormalities they should then consult a family doctor. In the third instance, the neurologist would come into play. But at the moment our healthcare system doesn't work like that. Today the doctor assesses the disease - and this is usually very subjective. We want to use machines to record disease symptoms more objectively.

The AI ​​is fed with data, which means it can also be subjectively colored. How objective can an AI actually be?

You're right. By “objective” I mean

less dependent on what an individual doctor sees and more patient-focused.” This is the potential benefit of AI.

Can you understand that many people are afraid of being judged by an algorithm?

I think this is an exaggerated fear. Many people are afraid that the algorithm will make a mistake. But just look at how many mistakes doctors make, who are often under enormous time pressure.

Would you already invest your money in one of the start-ups that promise to soon make a lot of money with voice analysis for disease detection?

Although I know companies that are active in this area, I have not invested in them. But I am also very conservative when it comes to investing.