DAILY DAY: Kappa Santé analyzes information collected on social networks. Can we really talk about health data?
Dr STÉPHANE SCHÜCK: I founded Kappa Santé in 2003, focusing it on real-life data and observational studies [l’entreprise a été vendue en 2023 au groupe espagnol Apices mais le Dr Schück reste président, NDLR]. About ten years ago, we lacked data from patients. We had experience with social media and saw it as an interesting source of health data. It had to be demonstrated. As soon as we talk about social networks, there are immediately quite strong preconceptions: about the fact that there are only people who complain, that it is fake news… All of this is partly true, but a lot of patients seek peer-to-peer support there, without referring to medical authorities, and express things that they do not necessarily share in an extremely standardized medical environment. Often, the patient does not make a single post but several. We then reconstruct the entire care pathway based on their activity on social networks.
Is this really representative?
When you are an epidemiologist, ensuring representativeness in studies is a daily battle. Indeed, social networks only represent the people who express themselves there. But the numbers are such – often hundreds of thousands – that ultimately, the notion of representativeness becomes relative. When you analyze the publications of 100,000 people with diabetes, this represents a mass almost never reached in observational studies or therapeutic trials. We are also able to identify the age or at least the age groups based on what is written.
We look at social networks from around the world, which allows us to do comparative analyzes between countries
What sources are analyzed?
Facebook, X, Instagram, TikTok, YouTube, Amazon… Nous screenons thousands of discussion forums and messages, but only public data that anyone can look at. We do not go to closed networks like the pages of patient associations. We draw from the famous Doctissimo, but there are other surprising sources such as Jeuxvideo.com, originally a forum for enthusiasts which today hosts hundreds of discussion topics, including health. We look at social networks from around the world, which allows us to do comparative analyzes between countries. For example, we studied the perception of patients and healthcare professionals about bariatric surgery in France and the United States.
What is your method?
The first challenge was to absorb these quantities of information which are renewed from minute to minute. We worked with a branch of artificial intelligence, NLP or automatic language processing. One of the difficulties lies in the language of patients, which is not the same as that of the medical profession. We spent a lot of time translating what patients were saying and coding it in MedDra [Dictionnaire médical pour les activités réglementées]. For example, someone with high blood pressure will not necessarily use these terms but will say that their “head is in a vice”.
Next, we developed algorithms exploring the topics we were interested in, such as quality of life and side effects. With the ANSM, we worked on Levothyrox to identify the signals when changing the formula. [en 2017, de nombreux patients avaient déclaré des effets indésirables à la suite d’une modification du traitement].
Social networks also allow you to travel in time! We can go back to 2009 and look at changes in the social outlook on health policies, the marketing of drugs, epidemics, etc. We also see the difficulties in accessing care and wandering about certain pathologies. We have worked on endometriosis and rare pathologies, where the diagnosis occurs years after the appearance of the first symptoms. Detecting these signals is good, but you still need ears to hear them.
Exactly, who do you work for?
We are a service provider. For example, we work for laboratories on real-life studies, on quality of life, side effects, but also the difficulties encountered in accessing treatments and the needs not covered in a pathology. We create dashboards for the detection of pharmacovigilance signals and analyze misuse.
We also looked at the detection of epidemics, particularly respiratory ones, as part of a post-Covid consortium funded by the European Hera Agency. The aim was to identify the occurrence of an episode early via social networks and correlate it with a signal for hospital emergencies.
Finally, we created a subsidiary Ultima-I to make predictions on clinical data using artificial intelligence. We will predict a patient’s response to a treatment, the side effects that might occur. We trained the algorithm using clinical data from social networks, clinical trials, but also administrative databases such as SNDS. [système national des données de santé]. Tomorrow, the healthcare professional will be able to query our tool by entering the characteristics of their patient.
