In 2023, more than one out of two French people said they had experienced mental suffering over the past 12 months. When the former French Prime Minister Gabriel Attal delivered his first general policy speech on January 30th, he stated that he wanted to make mental health, and particularly that of young people, a major focus of government action. The need to deploy an effective prevention and care system has therefore never been greater.
For about ten years now, Philippe Lenca, researcher at IMT Atlantique, has been working in cooperation with Brest University Hospital in the field of psychiatry. Along with his colleagues Romain Billot, Yannis Haralambous and, more recently, Gabor Bella, he has helped develop information and decision support systems for people with mental disorders. These systems can detect the early signs of serious mental health conditions such as schizophrenia and prevent recurring suicidal acts by analyzing patient databases.
The sociodemographic and medical datasets are enriched over time and feed risk models integrated into business applications such as VigilanS, a suicide risk prevention platform that has been implemented in several emergency centers in France, and coordinated with SAMU emergency services. These applications therefore provide decision-support tools for implementing appropriate actions such as maintaining contact with a person in distress, medical intervention, treatment and hospitalization.
Medical staff ensuring data quality
Establishing the foundations of these tools, the databases, is a tedious task in itself. Collecting high-quality sociodemographic and clinical information is an essential step in the early detection of mental disorders. However, it is dependent on the meticulous recording of care pathways (e.g. transcriptions of interviews, imaging, biology) within the health system, which does not always happen, especially in the case of emergency services.
To ensure the robustness of their data analysis, IMT Atlantique researchers only keep the pathways from people – diagnosed or at risk – which have been completed at least 80%. This drastically reduces the size of their samples. “This is unfortunate because in our analyses we usually look for the weak signals of the conditions,” Yanis Haralambous says. “We hope that this will help medical staff realize that, in order to be used, this data must be of good quality and must therefore be complete,” Philippe Lenca adds.
The quality of the data also depends on how the information is populated. “A few years ago, in the event of a suicide attempt, caregivers simply coded 1 or 0 depending on whether it was a recurrence of not, but not the number of times it had occurred. However, since suicide is usually preceded by several attempts; information is essential information for our prevention tools ,” Philippe Lenca says.
The slow creation of a patient cohort
Identifying an at-risk population is also challenging. In the case of schizophrenia, “it is not a rare disease in the medical sense of the term, yet it occurs infrequently. There are far fewer potential patients than for suicide [see box below],” Philippe Lenca explains. In Brest, out of a population pool of about 300,000 inhabitants, between 30 and 50 individuals at risk of developing a schizophrenic disorder are referred to the hospital each year by their general practitioner or social worker. Creating a cohort of at-risk patients is therefore a slow process.
“This referral is usually based on signs that are not clearly linked to depression or delirium. It often starts when the referring healthcare professional suspects something,” says Michel Walter, head of the psychiatry department of Brest University Hospital. These “suspected” individuals are then monitored for two years using assessment tools. Only 30% will then be diagnosed with a schizophrenic disorder, thus reducing the final sample to about ten individuals.
Towards multimodal data collection
In addition to the robustness of the data, the more extensive the databases, the better the tools. In order to improve its effectiveness, IMT Atlantique team plans to enrich the databases with additional clinical data. One of the avenues being explored is that of using linguistic data collected from psychiatric assessment interviews with automatic natural language processing techniques. The physiological signals recorded over the course of the care pathway can also be added to these datasets.
For the early detection of schizophrenia, for example, the research teams are considering using oculometric measurements. This data includes pupil size information, but also eye-tracking that monitors a person’s gaze. Patients at risk of schizophrenia have a distinctive eye path, with saccadic movements, including difficulties in following a moving object with their gaze. These eye movements are considered to be endophenotypes: biomarkers of vulnerability to the condition. The addition of eye-tracking techniques to the system would therefore make it possible to identify weak signals that are associated with the risk of developing the disorder.
The teams have already set up trials with patients to record their physiological data at home using a smartphone or a connected watch for example. This methodology opens up new opportunities in terms of data collection, including heart rate, sweating and sleep quality. Analysis of textual data extracted from social media is also being considered. Yet the possibility of using this method for early detection raises some ethical questions. What means of intervention can legitimately be put in place when a tendency is detected but does not yet exist?
we could be required to respond to a distress situation anywhere, including in the office. This does not only happen in private.
