More efficient and faster “screening” of burn-outs thanks to AI?


After a two-year health crisis, 41% of employees said they suffered from mental health problems due to lack of consistent protocols, changes in work speed or lack of possibilities. According to the latest OpinionWay survey for Human Footprint [cabinet spécialisé dans les risques psychosociaux et la qualité de vie au travail], 34% of employees are on burn-out, of which 13% are severe burn-out, i.e. 2.5 million people. According to the Health Watch Institute, 480,000 people in France have mental health problems at work and 7% or 30,000 burnout is alarming.

What are we talking about when we talk about burnout?

According to the INRS, burnout is characterized by severe mental fatigue and professional “feelings of inadequacy”. It is an extra work that drains the employee of all his energy. According to the WHO, it is a “pathology caused by chronic occupational stress that has not been properly managed” – including three components: “feelings of loss or fatigue”, “increased emotional distance from work or feelings of negativity or slander related to work”, and “work. Decreased effectiveness “.

To avoid a burnout or to get out of it, it is essential to be able to detect it as soon as possible; Especially through various warning signals, mental or physical. But health professionals are still struggling to diagnose the syndrome early. First, as professional trainer and psychosocial risk taker Emmanuel Wyert writes, “Burn-out: it’s not your fault, but it may be your fate”, because it is about a pathology consisting of “a rich and multifaceted symptomatology.” Thus, many of the symptoms may be manifestations of other illnesses or syndromes (depression or anxiety in others), and “do not burnout on their own merits”. Then, because the current diagnostic tools are “incomplete”.

Diagnostic equipment is still not reliable enough

The existing tools are far from benefiting from medical consent. The widely used MBI (Maslach Burnout Inventory), OLBI (Oldenburg Burnout Inventory), CBI (Copenhagen Burnout Inventory), SMBM (Shirom Melamed Burnout Measure) and BM (Burn-Out Measure) tests have dozens of questions to explore. Mental fatigue, relationships with others, and levels of achievement at work. They aim to assess the severity of the syndrome. Examples of standard questions and answers: “I feel tired at the end of my work day: never / never / every day”; Or even “it takes a lot of hard work to work with people all day: sometimes / sometimes / every day” “.

But if these experiments are verified by companies in psychology, “a lot of criticism is addressed to them, and they do not agree in the scientific community”, observes Emmanuel Wyrt. Especially since they are not reliable, the results are often biased “These questions contain only (closed) scale questions, and no free-text questions. In addition, some people are reluctant to tick the answer ‘never’ or ‘every day’, or are tempted to lie to influence the results,” Masha explained. Kurpiz-Bricki is a professor of data engineering at the Bernice University of Applied Sciences in Biel, Switzerland.

Together with AI and a team of applied psychology experts, this Swiss researcher developed a method based on automated text analysis to overcome this limitation. “Text analysis, for example in the form of interview transcripts, has been shown to hold promise, but is often not done in practice due to the extra time required for manual text evaluation. But with AI, we can detect burnout more easily and quickly. With reliability, ”he notes.

More efficient and faster “screening” of burn-outs thanks to AI?

The texts analyzed (anonymously) by Swiss researchers came from the English-language community website Reddit and its many forums or “subbradits” (and not just dedicated to burn-out). Mascha Kurpicz-Briki’s team thus created a dataset consisting of 13,568 samples describing first-hand experiences, including 352 burnouts and 979 frustration-related. He then resorted to machine learning. Specifically, algorithms automatically analyze texts and detect whether the language has been burned out. For this reason, the collected texts were labeled and categorized into two groups, among those referring to this syndrome and those not related to it. The models were then trained through a variety of configurations to determine if there was a “burn-out indication” in a text.

According to Swiss scientists, this method of diagnosis by natural language processing (NLP, automated natural language processing) can successfully detect 93% of burnout cases. But the number of data used to ensure complete results is still low. “As a next step, we want to apply our method in collaboration with clinical partners to further develop and validate such practices. In addition, we will need more diversified training datasets with different groups of the population, as we have so far worked with completely anonymous data.” Due to the difficulty of finding sufficient data, we used English texts in this study. Our medium term goal is to apply our methods to texts in German and French, from patients whose burnout is confirmed, “says Mascha Kurpicz-Briki.

For the researcher, the technology can be used in a clinical context, to “create new tools” for psychology / psychotherapy: “for example, interview transcripts can be analyzed, or new inventory forms can be defined and verified, including free text questions.” Using “. Can we go so far as to imagine doing the same thing by analyzing words verbally through the patient’s voice? “We’re exploring the possibility of auto-copying spoken texts, for example from interviews with patients, and then analyzing them. However, this can be difficult in certain dialects, which can be complicated to copy,” he says.

An “enhanced intelligence” without being artificial

But beware: there is no question of automating everything. Mascha Kurpicz-Briki points out that the collaboration of AI specialists and medical experts is essential. Initially to verify the findings of this study on the actual case of burnout and on a representative sample of the population, but also beyond. “Our work is based on enhanced intelligence approaches rather than artificial intelligence: instead of replacing humans (health professionals), such a tool should support them in their day-to-day work. Like, ”he explains.

Because researchers are not afraid to admit it: he and his team are unable to know exactly which word or phrase curves are being held by their AI system as a sign of burnout. Thus it is a method comparable to a “black box”, which is better not to use without minimal vision. “It is very difficult to assess how AI has come to such a conclusion. This is a challenge that we are trying to meet, but on full terms, this tool cannot be automated,” he concluded.

This type of technology can make “Sisi Chatbot” more “intelligent” which has been on the rise for 4 or 5 years. Owlie, Woebot, Wysa and Tess are the agents of the “Psychological Assistance” conversation, who ask you questions and give you advice to understand your moral and emotional state. Or refer you to a human specialist. These chatbots already detect keywords like “stress”, “insomnia”, “anger” via AI. Machine learning and natural language processing allow them to “help users manage their emotions” and “relieve their frustration”. But currently their effectiveness is limited. Hence the interest in the method created by Mascha Kurpiz-Bricki and his team. A tool that will certainly not replace human contractions, but will allow them to connect with their patients as upstream as possible.

The same scheme will apply to a chatbot that will verbally detect human emotions, as has been done. Eli. This This is how the “artificial incarnation” designed by researchers in computer science and psychology at the USC (University of Southern California) Institute of Creative Technologies. Able to communicate with patients, but able to identify their emotions Analyzes their facial expressions and sounds. Much like the Owlie or Woebot chatbot, Ellie’s AI starts with general questions and then moves on to more specific clinical questions.

Throughout the discussion, Eli uses the machine’s approach to interpreting oral and facial signals from the patient’s facial expressions and voice, to adapt his questions, but also his responses – which can nod, smile, or utter “hmm” when the subject is a sensitive story. Saying. These subtle “responses” thus make it possible to build a close “relationship” with the patient and invite him to share more information. If Mascha Kurpicz-Briki’s strategy can be applied to oral discussions, it is difficult to imagine its contribution to projects such as the USC. But here too people will be at the center. Gali Lucas, the researcher behind Eli, explains that AI can only be a service tool for human therapists: “With our virtual avatars, we don’t want to replace humans. Robots can detect better than our non-verbal information. This technology has great potential. Can be used to detect, but the ultimate care will always be with human therapists. “

Note that in Walloon Belgium, other researchers are working on a solution using AI to more easily detect burnout. Not based on text, but based on data analysis of patient’s smartphone and connected objects. “By observing behavioral data produced by phones and other connected devices, we can easily detect and detect key symptoms such as behavioral changes, physical fatigue and even possible abnormal emotional reactions,” explains Antoine Sepulcher, founder of NOÖS, a NOÖS platform that Dedicated 6 Mental Health and Wellness, who are collaborating on this project with experts from UCLouvain Artificial Intelligence Research Laboratory.

Behaviors analyzed through connected objects and other smartphones include changes in user’s marketing style, his voice tone, his sleep pattern, his search on Google, his eating habits, or his music preferences. A potentially effective method, but quite risky in terms of protecting personal (health) data and even ethically … That’s why, here again, researchers emphasize the importance of “putting people at the center. The solution.”

Leave a Comment