AlstroSight enters partnership with hospital practitioners

The decision to base themselves at Lyon Est Neurological Hospital was driven by both practical and scientific motivations for the multidisciplinary team AIstroSight, which specialises in using applied mathematics and computer science to understand neurological diseases and how the brain works more generally

As Maëlle Moranges explains, “We wanted to get closer to hospital practitioners and neuroscientists. Having this proximity not only to HCL, but also to specialist neuroscience and medical imaging laboratories such as the CRNL, the CERMEP and the SRBI, opens up new opportunities for research.

This also gives the team access to health data without the need to take it outside of the hospital sphere, thereby ensuring confidentiality and cybersecurity. AlstroSight employs the use of a wide range of biomedical data taken from sources such as cell culture or MRI imaging in order to create a homogeneous database capable of being operated using algorithms. 

But Maëlle Moranges ran into a major stumbling block. Despite its great potential in health, whether to improve diagnoses, personalise treatment or reduce medical errors, many medical practitioners remain unconvinced by artificial intelligence. “In a recent study carried out by researchers in the US, an AI system was found to be markedly better than a doctor at making diagnoses”, explains the postdoctoral researcher. “On its own it had a 92% success rate, while the doctor only had a 74% success rate. Working together their success rate was only 76%. The reason for this is that doctors are still reluctant to turn to AI.” 

How can health professionals be encouraged to buy in to AI? Maëlle Moranges thinks she has the answer. If doctors are to adopt AI and trust it, they need to understand how it works. This is the aim behind explainable AI, which Moranges is seeking to use. It provides a clear justification for doctors to accompany diagnoses made using AI.

Doctors find current explanations either overly mathematical or overly simplistic. Our goal is to develop AI systems that are capable of providing information that is intelligible but not too basic and adapted to the individual in question, whether they are a clinician, a researcher, a manufacturer or even a regulatory body.”

It’s a complex challenge. This will involve creating tools that are global enough to understand the general workings of the model but also local enough to justify decisions specific to individual patients. Herein lies the benefit of designing such a system with the doctors themselves.

This cooperation is very much central to the project being carried out by Maëlle Moranges in conjunction with the HCL’s Department of Vascular Neurology, where Professor Laura Mechtouff is assistant head. The project is seeking to explore ischaemic strokes, focusing on thrombectomy, a procedure that involves removing blood clots using suction or by inserting a stent into the obstructed artery. This is recommended up to 24 hours after the first symptoms of a stroke, either in addition to thrombolysis (using drugs to destroy the blood clot) or on its own.

But although this procedure is effective, it does have its limitations, as Maëlle Moranges explains: “One in two patients fail to recover their autonomy within three months of a thrombectomy. Laura Mechtouff believes that this could be down to excessive inflammation whereby the immune system aggravates the damage caused to a brain already weakened by the stroke.”

Seeking to test this idea, Maëlle Moranges wants to use AI to predict the risk of handicap for patients based on their inflammation profile. 

The long-term goal is to use large language models (LLM), machine learning models and explainability models to identify the blood biomarkers involved and then use these results as a basis for determining possible treatments to reduce inflammation.

The postdoc also has her sights set on another challenge put forward by Laura Mechtouff: why do women not recover as well as men from strokes? Could this be linked to an inflammatory reaction that varies depending on gender?

As we have seen, the importance of Moranges working directly with practitioners is clear. “Without this immersion in Laura Mechtouff’s department, we would never have considered inflammation as a possible cause of handicap.”

At a broader level, AlstroSight’s residency at HCL can be seen as an important step towards unlocking the potential of artificial intelligence in health, with artificial intelligence itself also benefiting in return. 

Experiencing day-to-day life in a hospital means we concentrate on subjects that are more useful to practitioners, linked to their concerns and with genuine clinical benefits. We explore avenues that haven't yet been tackled in the scientific literature, which haven’t been explained by conventional medicine and which are difficult to identify from a remote laboratory.

Another benefit is that when practitioners have access to vast quantities of data, it is difficult for them to make sense of it and identify what is having an impact. This is where algorithms step in, tireless explorers capable of sifting through these masses of information on their own. Seeking to develop explainability, Moranges turned to Pattern Mining which automatically detects recurrences and establishes links between pieces of data in order to describe conditions or highlight at-risk profiles. 

“If AI is able to combine both the descriptive and the prescriptive, it could become invaluable when it comes to making diagnoses and decision support.One final point is that medicine challenges us with highly complex problems, pushing us further when it comes to innovation. There is often a high degree of variability when it comes people's bodies and diseases, both at an intra- and an interindividual level, making them harder to predict than the purchasing habits of consumers, for example, an area in which AI is now widely used.”