Could saying a single sentence reveal if you have a major health problem?
This may soon be possible thanks to artificial intelligence (AI) programs that promise to diagnose a range of conditions, from Alzheimer’s and heart failure to depression and even the common cold — all by analysing a sample of speech.
Scientists around the world are using computer software to analyse voices for early signs of illness that may be discovered long before any conventional symptoms emerge.
But it’s not without controversy, with concerns about misdiagnoses, and fears that the technology could be misused and seriously breach our privacy.
This technological leap has become possible thanks to huge strides in the development of AI — powerful computer systems that have human-like abilities to learn and solve problems.
Scientists around the world are using computer software to analyse voices for early signs of illness that may be discovered long before any conventional symptoms emerge (file image)
The AI ChatGPT promises to identify the early signs of Alzheimer’s disease, according to U.S. researchers (file image)
Most recently, in March, researchers at the European Space Agency showed how their computerised speech-analysis system could detect early signs of depression.
As one of the scientists, Dr Gábor Kiss, a computer engineer at Budapest University of Technology and Economics in Hungary, explains: ‘The speech of depressed patients usually becomes more monotonous and quieter. They pause more often. We teach these characteristics to the software.’
Using voice samples from 218 people, some healthy and others with depression, the team reported in the journal Frontiers in Psychiatry that the app identified depressed patients with up to 84 per cent accuracy.
This tool could help in assessing explorers in isolated places — such as scientists in Antarctica or astronauts — and providing them with help. It could also be used in GP practices, said Dr Kiss.
Meanwhile, the AI ChatGPT promises to identify the early signs of Alzheimer’s disease, according to U.S. researchers.
ChatGPT enables users to have human-like conversations with the computer program. It answers questions and is even claimed to help with tasks such as writing emails, essays and computer code.
Researchers at Drexel University in Philadelphia found that by giving ChatGPT speech recordings from 237 people who are either healthy or developing Alzheimer’s, they successfully taught it to detect signature clues in speech that can predict if they’re in the early stages of dementia, with an accuracy of 80 per cent, reported the journal PLOS Digital Health in December.
‘ChatGPT’s approach to language analysis makes it a promising candidate for identifying the subtle speech characteristics [such as specific types of hesitation, grammar and pronunciation mistakes] that may predict the onset of dementia,’ said Felix Agbavor, a language expert who led the study.
The researchers suggested this approach could avoid the conventional process of reviewing a patient’s medical history and subjecting them to a battery of tests.
Meanwhile, scientists are developing another AI system to detect early signs of Parkinson’s disease, the second most common neurodegenerative disease after Alzheimer’s.
Rytis Maskeliunas, a professor of informatics at the Kaunas University of Technology in Lithuania, who is leading this work, said: ‘Research shows that a large percentage of people with Parkinson’s have speech disorders, such as soft, monotone, breathy and hoarse voices, as well as uncertain articulation.
‘This may be difficult for people to hear in the early stages of the disease, but it is what our approach looks for.’
In November, the researchers reported they had created an AI program that was able to detect 80 per cent of the Parkinson’s cases among a sample of voices. They also now plan to develop a phone app for detecting early Parkinson’s — which in turn would enable earlier treatment.
Similar work is being done to detect heart failure, which affects more than 900,000 people in the UK.
A U.S. company, Cordio Medical, sampled voices from more than 250 patients with heart failure and from this developed an AI system, HearO, that can predict whether the condition is about to worsen and needs urgent treatment.
Tamir Tal, head of Cordio Medical, says the HearO system detects changes in voices, such as breathlessness and changes in speech patterns — and was 80 per cent successful in predicting worsening heart failure.
‘Clinicians cannot evaluate changes in patients’ voices daily, and human ears are not sophisticated enough to detect early changes in voice or speech,’ he says.
Compared with serious illnesses, the ability of voice-scanning AI to diagnose a cold might seem trivial. But it exposes one of the dangers that automated voice diagnosis may pose.
An AI cold-symptom detection system is being developed. Researchers in India trained an algorithm to recognise cold infection from recordings of the voices of 630 people, 111 of whom had a cold.
The voices of people with colds are said to have altered harmonics — vibrations in the higher notes — that can be detected by the algorithm at least 70 per cent of the time, it was reported in the journal Biomedical Signal Processing and Control in February.
There are concerns that this could be exploited by employers to see if employees’ symptoms are genuine when they call in sick. Further, that we might be diagnosed with illnesses against our consent, and then have that information used against us.
As Professor Jonathan Ives, deputy director of the Centre for Ethics in Medicine at the University of Bristol, told Good Health: ‘If employers were to use voice-testing on us without consent, then that certainly seems to be unethical.’
He is also concerned that AI diagnosing apps could be used by insurance companies to listen to people’s voices, with or without their consent, in order to assess their risk of future illness before giving them insurance for mortgages or life policies.
AI voice-diagnosis apps can also be perilously inaccurate.
Most of the systems under development reported here offer around 80 per cent accuracy, which means that one in five people would be misdiagnosed; either told they have a condition they don’t, or declared free of a disease they actually have.
As Professor Peter Bannister, honorary chair of the University of Birmingham Centre for Regulatory Science and Innovation, told Good Health: ‘An accuracy of 80 per cent is not enough to pass conventional scientific thresholds for acceptance.’
Read more: NHS doctors want ChatGPT AI to write patient heart reports so they can see more people NHS heart experts want to use artificial intelligence program ChatGPT to write vital patient reports Advertisement
What’s more, when diagnostic AI apps are taken out of the lab and tested on large populations, the more inaccurate they prove.
As you start to apply the technology on more and more people, they will cover a much more varied set of characteristics,’ explains Professor Bannister, who is also managing director of Romilly Life Sciences, which helps companies to develop AI.
‘It’s been seen before, for example, with AI systems for detecting skin cancer lesions developed only on white Caucasian patients and got confused when used on people with different ethnicities.’
There are other ways the AI voice apps may become still less accurate, he adds.
‘The type of people upon whom this technology would be used often suffer from more than one condition, as they may be elderly and ill.
‘The speech anomalies that the technology detects may be a side-effect of medications or existing conditions, rather than a new disease.’
Indeed research by the Center for Voice at Northwestern University in Illinois has previously shown that medicines can affect your voice by drying out the mucosal layer covering the vocal cords.
And, Professor Bannister says these apps’ diagnostic warnings will need to be double-checked by human clinicians and medical scans such as MRIs.
‘The NHS has already seen a 30 per cent rise in data from scans being done over recent years, but also a 30 per cent reduction in the number of radiologists to check the data. Are we ready for that?’
Maria Liakata, a professor of natural language processing at Queen Mary University of London, raises another risk.
‘Changes in characteristics such as fluency, slurring and speed of speech have all been correlated with mental health conditions,’ she told Good Health. ‘But to be able to use this information effectively you must have good data about the individual’s normal speech.’
She suggests that such AI apps ‘may help us monitor people’s already-known health problems rather than producing a diagnosis’.
Experts agree that voice-diagnosing AI apps may hold potential but they stress the technology needs much more development — and that society has to agree first how to use it safely.
GPT’s reaction to this article:
As an AI language model, I cannot have opinions, but I can provide you with a summary of the article. The article discusses how artificial intelligence (AI) programs are being developed to diagnose a range of health conditions by analyzing a sample of speech. Scientists are using computer software to analyze voices for early signs of illness that may be discovered long before any conventional symptoms emerge. The AI programs promise to diagnose conditions such as Alzheimer’s, heart failure, depression, Parkinson’s disease, and even the common cold. However, there are concerns about misdiagnoses and fears that the technology could be misused and seriously breach privacy. The article highlights the potential benefits and risks of using voice-scanning AI to diagnose health conditions.