Health: the development of the diagnosis by voice |Brain & Psycho

In March 2020, when it became clear that the coronavirus pandemic took on an unprecedented scale, officials from around the world began to ask everyone to participate in the fight.Hospitals have encouraged local businesses to give protective masks.The researchers enjoined those who had recovered from the COVVI-19 to give their blood plasma.And in Israel, the Ministry of Defense and a young company called Vocalis Health, established in this country and in the United States, asked the inhabitants to give...their voice.

Specializing in the analysis of vocal characteristics, Vocalis had already created an application for smartphone which detects the chronic obstructive bronchopneumopathy by analyzing the signs of breath of users when they speak.The company wanted to develop the same type of device for the COVVI-19.People tested positive for coronavirus could participate in the study by downloading an application made available by vocalis for this research.Once a day, they launched the application and spoke in their phone, describing an image aloud and counting from 50 to 70.

Diagnose the COVID-19 thanks to the voice

Thanks to an automatic learning system (Machine Learning), Vocalis then processed these records, as well as the votes of people whose screening test was negative, to try to identify a vocal imprint of the disease.In the middle of summer, the company had more than 1,500 voice samples and the pilot version of a digital COVVI-19 screening tool.This tool, which the company currently tests worldwide, is not intended to provide a final diagnosis, but to help clinicians to sort potential cases: it is a question of identifying those who would most needbe tested, quarantined or neat by medical staff."Perhaps our AI algorithm will be useful from this point of view?"Wonders Tal Wenderow, President and Managing Director of Vocalis."This method is not invasive, it is not a drug, we do not change anything in the patient.All you have to do is talk.»»

This company is not the only one to have embarked on the race for vocal biomarkers of the COVID-19-at least three other research groups work on similar projects.Other teams use cough audio recordings caused by coronavirus or develop algorithms that detect if a person carries a facial mask from the analysis of his voice.

This shows how the young vocal diagnosis is impatient to make a place in the sun.Over the past decade, scientists have used artificial intelligence systems (AI) and automatic learning to identify potential vocal biomarkers for a wide variety of pathologies, such as dementia, depression, disorders of the autism autismand even heart disease.The technologies they have developed are capable of detecting subtle differences in the way of talking about patients, and businesses around the world are starting to market them.

« Siri, est-ce que je suis malade ?»»

For the moment, most teams adopt a slow and progressive approach, designing tailor -made tools for medical offices or clinical trials.But many dream of deploying this technology on a larger scale, using microphones that are omnipresent in the products we buy.These systems could one day allow epidemiologists to use smartphones to follow the spread of diseases, and transform smart speakers into home medical devices. « À l’avenir, votre robot, votre Siri, votre Alexa [des assistants vocaux respectivement développés par Apple et Amazon, ndlr] dira simplement : “Oh, vous avez un rhume”»», explique Björn Schuller, spécialiste de la reconnaissance de la parole et des émotions, qui occupe un poste conjoint à l’université d’Augsbourg, en Allemagne, et à l’Imperial College de Londres, et qui dirige l’une des études sur le Covid-19.

But automated vocal analysis still represents a new field, which comes up against a number of potential pitfalls, ranging from erroneous diagnostics to intrusions in privacy and in personal medical data.Many studies are still preliminary, on a small scale, and it will not be easy to reach a finished product. « Nous n’en sommes qu’au début»», déclare Schuller.

Certain conditions cause obvious vocal distortions - think for example of the diction of a person suffering from allergies and who is suddenly grasped from a feeling of suffocation.But many scientists believe that vocal analysis could go far beyond and help identify a very wide range of disorders, thanks to the complexity of human speech.

Speaking requires the coordination of many anatomical structures and systems.At the start, the lungs send air through the vocal cords.These produce sounds, which are then shaped by language, lips and nasal cavities, among other structures.The brain, with other parts of the nervous system, regulates all these processes and determines the words pronounced.A disease affecting one or the other of these structures can leave traces in speech, which are all clues for the diagnosis.

Thousands of samples of voice analyzed

Automatic learning offers a way to detect these anomalies, quickly and on a large scale.It is now possible to introduce hundreds or thousands of voice samples in a computer, which is looking for the characteristics specific to patients with various pathologies.

Most of the first works in this area focused on Parkinson's disease, which has well -known effects on speech and for which there is no test delivering a final diagnosis.This disease causes various motor symptoms, including tremors, muscle rigidity and problems of balance and coordination.The loss of control extends to the muscles involved in speech;As a result, many Parkinsonian patients have a weak and soft voice. « C’est un des éléments décelables par l’oreille humaine»», explique Reza Hosseini Ghomi, neuropsychiatre à l’EvergreenHealth de Kirkland, dans l’État de Washington, qui a identifié des caractéristiques vocales associées à plusieurs maladies neurodégénératives."But with 10,000 samples and a computer, you will be much more precise.»»

More than ten years ago, Max Little, researcher in automatic learning and signal processing, today at the University of Birmingham, in the United Kingdom, looked at the feasibility of voice analysis applicationsfor helping difficult diagnostics. Dans une étude, Little et ses collègues ont utilisé des enregistrements audio de 43 adultes, dont 33 patients parkinsoniens, prononçant la syllabe « A»» de manière prolongée.Thanks to speech treatment algorithms, they analyzed 132 acoustic characteristics of each recording, to finally retain 10 which seemed the most predictive of Parkinson's disease - like shortness of breath and trembling oscillations of height and stamp.Based only on these 10 characteristics, the system identified the speaking samples produced by patients with an accuracy of almost 99 %.

With other researchers, Little has also shown that certain vocal characteristics are linked to the severity of the symptoms of Parkinson's disease.According to him, the systems are not yet robust enough to be used routine in clinical practice, but there are many potential applications.Vocal analysis could constitute a quick and inexpensive way to monitor people at risk, detect large populations or even create a telephone service to distant remotely those who do not have access to a neurologist.A portable version of the device - in the form of a smartphone application, for example - would also allow patients to follow their own symptoms and monitor their response to drugs. « Ce type de technologie est capable de fournir des instantanés à haute vitesse, presque un suivi continu, de l’évolution des symptômes»», explique Little.

From Parkinson to Alzheimer

Researchers are now trying to identify word -based biomarkers for other types of neurodegenerative diseases. En analysant des échantillons de voix et des transcriptions de parole fournis par plus de 250 personnes, un trio de scientifiques canadiens a par exemple repéré des dizaines de différences entre celles qui avaient reçu un diagnostic de maladie d’Alzheimer « possible ou probable»» et les autres.Participants affected by pathology thus tended to use shorter words, a more limited vocabulary and more sentences of sentences. Ils se répétaient en outre plus souvent et utilisaient davantage de pronoms, comme « il»» ou « ceci»». « Cela peut être un signe qu’ils ne se souviennent tout simplement pas du nom des choses et qu’ils doivent se servir de pronoms à la place»», explique Frank Rudzicz, informaticien à l’université de Toronto, qui a dirigé l’étude.

How to write poem | कविता कैसे लिखे What makes a bunch of words to poem. Part 1 https://t.co/j2hTI0xbrD #KavitaKaAalay #Hindipoetry

— काव्यालय Sun May 05 04:15:47 +0000 2019

Taking into account 35 of these vocal characteristics, the system has succeeded in identifying people with Alzheimer's disease with an accuracy of 82 % (this rate has since improved to reach around 92 %, specifies Rudzicz, notingthat errors are distributed more or less also between false negative and false positive)."These characteristics add up to form a sort of fingerprint of dementia," explains the scientist.This is a very complex hidden pattern that we have trouble identifying, but that automatic learning is able to identify, provided that you have enough data.»»

Blood test or sound recording?

Since some of these vocal changes occur from the early stages of neurodegenerative diseases, researchers hope that voice analysis tools will one day help clinicians diagnose these pathologies earlier and intervene before other symptoms n'appear. Pour l’instant, cette idée reste toutefois largement théorique et son potentiel doit être confirmé par des essais dits « longitudinaux»», qui portent sur un grand nombre de patients suivis sur le long terme.Some doctors also point out that the analysis of the voice alone will rarely make it possible to make final diagnoses. « J’apprends beaucoup en écoutant la voix de quelqu’un»», déclare Norman Hogikyan, laryngologue à l’université du Michigan, à Ann Arbor." It is my job.But I combine it with an analysis of the patient's medical history and then an examination.The three parts of this evaluation are important.»»

Active researchers in this area point out that the objective is not to replace doctors or create autonomous diagnostic devices. Ils voient plutôt l’analyse de la voix comme un outil d’aide à la décision pour les soignants : ce serait un « signe vital»» supplémentaire à surveiller ou un test qu’il est possible de demander. « Je pense que la collecte d’échantillons de parole deviendra aussi courante qu’une prise de sang»», déclare Isabel Trancoso, spécialiste du langage parlé à l’université de Lisbonne.

A certain number of young companies specializing in vocal analysis - including Winterlight Labs, a Toronto company co -founded by Frank Rudzicz, and AURAL Analytics, in Scottsdale, in Arizona - now provide their software to pharmaceutical companies.Many of them use it to determine whether their experimental treatments act on people registered in clinical trials. « En utilisant la parole comme un indicateur plus subtil des changements dans la santé neurologique, on peut aider les médicaments à franchir la ligne d’arrivée ou, à tout le moins, identifier rapidement ceux qui ne sont pas prometteurs»», explique Visar Berisha, directeur des recherches et cofondateur d’Aural Analytics.

Voice signs of autism from the age of ten months?

Neurodegenerative diseases are only the beginning.Scientists have also identified specific language patterns in children with neurodevelopmental disorders.In a 2017 small -scale study, Björn Schuller and his colleagues showed that the analysis of the bailling of 10 months old infants thanks to algorithms made it possible to identify with a certain precision those who would later be diagnosed as suffering from spectrum disordersautistic. Le système a correctement classé environ 80 % des enfants autistes et 70 % des enfants « neurotypiques»» [une appellation utilisée pour désigner les enfants ne souffrant pas de troubles du spectre autistique, ndlr].

Les chercheurs ont également constaté que de nombreux enfants souffrant de trouble déficitaire de l’attention avec hyperactivité parlent plus fort et plus vite que leurs camarades, et quetheir voice présente davantage de signes de tension.Peakprofiling company in Berlin is currently developing a clinical speech analysis tool which, she hopes, will help doctors diagnose this disorder.

But some clinicians are skeptical about the amount of useful information that these systems are really able to bring. « Il y a une part d’exagération»», déclare Rhea Paul, spécialiste des troubles de la communication à l’université Sacred Heart de Fairfield, dans le Connecticut.Children with neurological development disorders often have many easily observable behavioral symptoms, she notes.

Ne pas « étiqueter»» les enfants trop tôt

In addition, it is not yet known whether algorithms really identify specific markers, for example for autism spectrum disorders, or if they simply detect general signs of atypical cerebral development - even temporary anomalies of speech. « Le développement est un chemin sinueux et tous les enfants qui développent des signes d’autisme ne deviennent pas des adultes autistes»», explique Rhea Paul.Even if scientists manage to identify a specific and very reliable vocal biomarker, she adds, it should only be used to identify cases where a more in-depth evaluation would be profitable."It should not be considered enough to label a child, especially so early in his life.»»

Vocal analysis technologies are also being studied in the context of mental illnesses.Many teams around the world have developed systems capable of detecting slow, monotonous and punctuated breaks that often characterizes depression, and others have found vocal biomarkers associated with psychosis, suicidal trends andbipolar disorders. « La voix véhicule une multitude de signaux émotionnels»», explique Charles Marmar, psychiatre à l’université de New York."Speed, rhythm, volume, height, prosody [accentuation and intonation, note] - these characteristics tell you if a patient is depressed and discouraged, if he is agitated and anxious, or S 'He is dysphoric [plunged into a state of depression, dissatisfaction, anxiety, editor's note] and maniac.»»

In his own works, carried out with 129 military veterans-all men-, Marmar used automatic learning to establish 18 vocal characteristics associated with post-traumatic stress syndrome (SSPT).These were mainly indicators of a slow, dull and monotonous speech.By analyzing these characteristics, the system detected veterans who suffered from post-traumatic stress with an accuracy of almost 90 %.

High speed screening of post-traumatic stress

Marmar and his colleagues are now trying to include women and civilians, in order to generalize their results.If they succeed, their technology would allow them to quickly identify people who need a more in -depth psychiatric assessment."The first concrete application would be high-speed screening for post-traumatic stress syndrome," explains Marmar.It is possible to perform 4,000 vocal screening in a few hours.»»

Similar consumer applications are already starting to emerge.One of them, intended for the American Ministry of Veterans, aims to follow the mental health of employees, in order to identify those who suffer from psychological distress.This smartphone application, developed by a company based in Boston and named Cogito, collects metadata on user habits - such as the frequency to which they call or send SMS to other people - and analyze the vocal memos theyleave on their phone.

Perhaps we will even find vocal biomarkers for pathologies that seem without any connection to speech.In 2018, scientists have analyzed speech samples of 101 people who were to suffer an examination of the arteries and blood vessels which fuel the heart.They then discovered that certain characteristics of vocal frequencies were associated with a more serious coronary disease.

We do not know exactly what explains these differences. « Nous avons du mal à comprendre le lien de cause à effet parce qu’il n’est pas évident»», explique Amir Lerman, cardiologue à la Mayo Clinic de Rochester, dans le Minnesota, qui a dirigé les recherches.Coronary disease could theoretically change the voice by reducing blood flow, according to him.But it is also possible that it is not the pathology itself that causes voice changes, but other associated risk factors, such as stress or depression.

The limits of the vocal diagnosis

This study shows both the promises and limits of this technology.It is one thing to detect computer vocal patterns, but it is another, more difficult, to understand what they mean and if they are of clinical importance.Do they translate fundamental characteristics of the disease?Or do they flow from another difference between groups, such as age, sex, size, education or fatigue?"We are trying not to be content to introduce figures in an algorithm," explains the neuropsychiatrist Reza Hosseini Ghomi.Our approach is rather to immerse ourselves in datasets, first develop a model of the disease, then to test it with automatic learning.»»

Until now, most studies have identified potential biomarkers in a single patient population, small, which is more."Reproducibility is still questioning," says Lerman. Est-ce que ma voix sera la même aujourd’hui, demain et après-demain ?»» Pour s’assurer que les résultats sont généralisables – et pour réduire le risque de biais, véritable plaie des algorithmes médicaux –, les chercheurs devront tester leurs systèmes de classification sur des échantillons plus grands, plus diversifiés et incluant des langues variées. « Nous ne voulons pas valider un modèle vocal avec seulement 300 patients»», explique Jim Schwoebel, responsable des données et de la recherche chez Sonde Health, une société d’analyse vocale basée à Boston."We think we need at least 10,000 patients, or more.»»

The company manages Surveylex, an online platform that allows researchers to easily launch surveys including vocal analyzes.She also takes care of the Voiceome project, which aims to collect voice samples and health information of 100,000 people, sweeping a wide variety of vocal tasks, places and accents. « Si vous êtes déprimé à New York, les modifications de votre voix qui s’ensuivent seront peut-être interprétées comme le signe d’une humeur différente à Houston, au Texas»», explique Schwoebel.

The difficult exit from the laboratory

For many of the applications envisaged by researchers, vocal analysis systems will have to distinguish not only between sick people and those who are healthy, but also between various disorders.And this outside the laboratory, under the more or less random conditions of daily life, and by exploiting recordings collected on a wide variety of general public devices. « Les smartphones disposent d’une gamme limitée de capteurs et les gens les utilisent partout, dans des environnements très peu contrôlés»», explique Julien Epps, qui étudie le traitement des signaux vocaux à l’université de Nouvelle-Galles du Sud, à Sydney, en Australie.

When EPPS and his colleagues, including a healthy proxy researcher, analyzed laboratory -registered voice samples with high -quality microphones, they detected depressive patients with approximately 94 % precision.Based on vocal samples that participants had recorded on their own smartphone, in their living environment, precision fell at less than 75 %, the researchers said in a article published in 2019.

And it is not because this technology is not invasive that it is without risk.It thus poses serious confidentiality problems: the danger is in particular that speakers are identified from anonymous speech samples, that private conversations are inadvertently recorded by analysis systems and that sensitive medical information are hacked,Sold, shared, or used to unscatm.If technology is not properly regulated, insurers and employers may also use these systems to analyze speaking samples without the explicit consent of their interlocutor and to obtain information on their health.With the key to potential discrimination against their customers or their employees.

Without forgetting the permanent risk of false positives and overdiagnosis."We must be realistic and understand that all this is still largely in the field of research," says Rudzicz.And we have to start thinking about what will happen when we put it into practice.»»