Diagnostic assistance based on patient feedback
Marc Shawky is a professor of computer engineering and a member of the ‘Care, Vulnerability and Technologies’ team, formerly CRI, at the UTC Costech Laboratory. He is working on the development of tools to aid in the diagnosis of Lyme disease and decision-making based on patient feedback.
“The restructuring of our research team more clearly highlights the team’s areas of focus, emphasising the ’care” aspect of our approaches. Current work on Lyme disease is part of the Num4Lyme and Dialyme projects, funded respectively by the Sorbonne Centre for Artificial Intelligence (SCAI) of the Sorbonne University Alliance and the Hauts-de-France Region,” explains Marc Shawky.
The research focuses on big data analysis and machine learning, which poses challenges not only in terms of data collection but also data security.
The challenge of data collection? “Initially, we thought we would work with data from the Saint-Côme Polyclinic in Compiègne. However, not only did it take a long time to set up the agreement, but the data available was not fully digitised. To save time, we retrieved anonymised data from 450 patients at Mater Misericordiae Hospital in Dublin, which was looking for partners to evaluate antibiotic treatments. This database is very specific in that it consists of Lyme disease patients before and after treatment. Other data from the American ILADS network, which specialises in patients with vector-borne diseases, also supplemented our data,” he explains.
What about data security? “As UTC does not have the option of having personal data managed by the IT department, we opted for completely anonymised data. We also sought the opinion of the Sorbonne University Ethics Committee, to which we report, providing it with details of all the IT processing carried out on this data, as well as the audit by the UTC IT department, in order to reassure it about security aspects. This resulted in a favourable opinion,” he explains.
Marc Shawky: what is unique about your team approach? “At UTC-Costech, the analysis of this data is part of the synergy between the approaches of “providing care”, i.e., providing medical treatment, and “caring”. The question we asked ourselves in the first case was: what can AI contribute and what type of AI do we want to use? From the outset, we focused on developing models centred on patients’ feelings. In short, we wanted patients to be able to express their condition as they perceived it. We wanted symptoms and clinical signs to be an integral part of the analysis. However, these clinical signs remain subjective and can involve nearly 200 different parameters, in addition to the biochemical parameters from blood tests and other serological tests,” emphasises Marc Shawky.
These analyses can include nearly 400 specific parameters per patient. “For example, we may have 10 clinical signs relating to joint pain or headaches. They can also be cognitive in nature, such as memory disorders, particularly short-term memory. Other symptoms such as cardiac arrhythmia or extreme fatigue are often cited, to which other signs may be added in the event of co-infection. Indeed, when a tick bites, it can transmit Lyme disease, but also other bacteria, viruses or parasites,” he explains.
Data analysis therefore relies on AI and, more specifically, on machine learning, with approaches capable of combining subjective and clearly numerical parameters. “Initially, the idea was to develop a diagnostic tool, but we quickly realised that providing a positive or negative diagnosis was not effective. We shifted our focus to developing a decision-making aid for hospital doctors. A tool with a feature that calculates the proximity between patient profiles. Computer optimisations were developed to process data locally on the clinician’s computer, transferring only the software code from the server to the computer and not the patient data to the server,” he adds.
How does this play out, in concrete terms? “In a hospital setting, doctors will use this tool to select patients from the database whom they have decided to treat because they consider that they have sufficient evidence to prove that Lyme disease has been contracted. But this same tool will also enable them to calculate the proximity of the profiles of patients treated with other patients in the database, in order to determine the most similar profiles and issue recommendations that the practitioner may or may not follow,” concludes Marc Shawky.
MSD
