Better diagnose transplant rejection using artificial intelligence. So far the idea seemed futuristic. However, researchers from Paris Cité University and Inserm have just developed a “computer assistant” capable of making diagnoses of kidney rejection more reliable. The results were published in the journal Nature Medicine. Having proven itself, the tool is already available for self-service for transplant centers in France and abroad. The researchers hope that it will gradually be adopted by hospital departments around the world.
“These are the first steps towards so-called precision medicine in which doctors provide better diagnoses using algorithmic tools. But beware, the purpose of this tool is to assist the doctor and not to replace him”, specifies Alexandre Loupy, nephrologist at the Necker hospital, director of the research team in organ transplantation at the Cardiovascular Research Center of Paris. (PARCC).
In France, just over 3,000 patients each year benefit from a kidney transplant, but 10 to 20% of them reject it in the months following the operation. Improving the diagnosis of rejection is therefore a major challenge. Rising to the challenge, Alexandre Loupy’s team developed a specialized automated computer assistant. And the machine spotted a 40% rate of imperfections in the assessments made by the doctors. What open new avenues for improvement in the care of patients.
” READ ALSO – Kidney failure: the obstacle of the shortage of grafts
Until now, the diagnosis of kidney transplant rejection was based on the Banff international classification, established in 1991. “This classification makes it possible to define whether the patient has transplant rejection or not based on clinical, immunological and histological data, c ie biopsies of the graft”, explains Professor Christophe Mariat, nephrologist at the Saint-Étienne University Hospital and president of the transplantation commission of the French-speaking Society of Nephrology, Dialysis and Transplantation. “But the main limitation of this classification is its lack of reproducibility.” The new algorithm has, however, made it possible to optimize these criteria and it works in a totally transparent way by indicating each step that led to the decision of “rejection” or “no rejection”, the whole taking the form of a “tree decision-making”.
In order for the algorithm to work, the scientists fed it for three years. Two additional years allowed the performance of the tool to be evaluated by comparing the diagnoses proposed by the machine with those of doctors. These comparisons involved a total of 4409 biopsies from 3054 transplant patients in Europe and North America. “We were surprised to see such an amplitude of inaccuracy, ie 40%, in the doctors’ assessments”, says Professor Loupy. However, an imperfect diagnosis can have consequences for transplant recipients: “a person misdiagnosed is likely to receive immunosuppressive treatment (anti-rejection, editor’s note) wrongly. And these heavy treatments can be accompanied by side effects. Conversely, we can miss a genuine rejection, which leads to the under-treatment of the patients concerned”, underlines Professor Mariat.
Reducing these diagnostic errors is the researchers’ ultimate objective, but this consequence will be indirect. “Initially, it will be a question of categorizing the patients according to the nature of their rejection and then of evaluating the effectiveness of therapeutic treatments on these homogeneous populations of patients in order to determine which is the most effective for this or that another type of rejection. “says Alexandre Loupy.
” READ ALSO – Organ transplants: a worrying decline for five years
The computer assistant has already been introduced to international transplant societies. “The French are still reluctant to use it, while the international experts responsible for updating the international classification have already approved the system”, points out Alexandre Loupy. Doctors at Necker Hospital in Paris and Cedars-Sinai Medical Center in Los Angeles are already experimenting with it. The inventors hope that other hospitals will take ownership of the tool “so that we can all speak a common language”, enthuses Christophe Mariat.
However, the tool has its limits. The biopsy is always read by pathologists. The latter interpret the biopsy samples then enter their results into the machine, which synthesizes them and develops a diagnosis. “So there is always the risk of human error,” warns Alexandre Loupy. “A next step will be to develop an artificial intelligence capable of both reading the biopsy fragment, thanks to cameras, so that it can itself detect elementary lesions before making the diagnosis. We are already working on it.”