By akademiotoelektronik, 08/02/2022
How Algorithms Help Stem Cancer
Decryption#medtechReading time: 09'25''November 23, 2021[INNOVATING AGAINST CANCER 1/5] If cancer remains the second leading cause of death in the world, many startups are using artificial intelligence to identify it from all sides .LISTEN TO ARTICLESHAREREPORT AN ERRORSAVE PDF / EXPORTHow Algorithms Help Stem Cancer00:00 - 00:0000:00In This Article1. Improving prevention2. Accelerate and clarify the diagnosis3. Optimizing the choice of treatment4. Better prevent relapses
Will artificial intelligence succeed where humans have not yet been able to? Man knows how to send satellites into space, make peers walk on the Moon and explore Mars, but remains powerless in the face of cancer. Protean, these pathologies are the cause of around 10 million deaths per year, according to the World Health Organization (WHO), i.e. nearly one in six deaths, and constitute the second leading cause of death in the world. . Despite advances in research, some cancers remain extremely difficult to prevent, diagnose and treat.
Many startups have therefore set themselves the challenge of improving patient care, from the prevention phase to support during their treatment. And artificial intelligence ranks high among the most popular technologies for entrepreneurs to overcome this global scourge. Researchers and entrepreneurs unleash their increasingly powerful algorithms on the mountains of data available on the subject. Should this be seen as a miracle solution? What can computers really do against cancer? Overview.
Improving prevention
This is one of the major areas of research today: how to improve prevention to avoid the appearance of certain cancers? This is particularly the case for the most documented cancers, for which risk factors are known, such as cancer of the cervix or breast. The better these factors are known, the easier it is to implement the prevention strategy: in the case of cancer of the cervix, a vaccine against the papillomavirus responsible for the development of the disease has been developed; in breast cancer, a regular screening strategy makes it possible to anticipate treatment.
But, in many other cases, "we don't know why the cancer occurs", recalls Philippe-Jean Bousquet, head of the Observation, Data Sciences and Evaluation department at the National Cancer Institute. (INCa). The study of large amounts of data using algorithms therefore makes it possible to “characterize risk factors thanks to subgroups formed by AI, which are not necessarily logical, which a scientist would not have done”. “If the AI did it, it's interesting to ask how it characterized the profiles, it helps generate hypotheses. This is not an affirmation, but it allows us to ask new questions. And thus to identify new risk factors, hitherto passed under the radar of the human eye.
Secondly, the technology could be used to modify public screening policies, around the audiences concerned. "We can ask ourselves the question of the role that artificial intelligence could have in the organized screening of breast cancer", questions Philippe-Jean Bousquet. Once the risk factors have been identified, the screening strategy could indeed be adapted so that the efforts are mainly focused on the population presenting these factors. This is the idea behind the startup Predilife which, thanks to the study of blood and saliva samples by artificial intelligence, is able "to identify the main risks of breast, lung, prostate, colorectal cancer , melanoma as well as cardiovascular diseases”, assures the company. And therefore to sensitize individuals presenting the greatest risks to the importance of being closely monitored and regularly screened.
Accelerate and clarify the diagnosis
Today, this is the stage where artificial intelligence is used the most. Many startups train their algorithms to allow them to make an initial diagnosis and thus accelerate the detection of cancerous strains. "More than 250 million analyzes are carried out each year worldwide, including many early diagnoses", explains Fanny Sockeel, CEO of Primaa, a startup whose AI is capable of identifying certain types of cancer from samples. biological. “Some benign cancers can be detected early, others require a more complex diagnosis, take a long time to be characterized from the pharmacogenomic profiles of patients. It is to allow doctors to focus on these complex cases that Primaa's AI is a diagnostic aid in the simplest cases. Other startups, like Magic-Lemp or Witsee, offer the same type of service.
Developing this expertise is far from easy. Artificial intelligence must be able to "manage the specificities specific to each analysis laboratory: each has its own techniques for preparing or staining samples, glass or plastic slides, scanners with a particular leg which means that the images are not all alike”, explains Fanny Sockeel. However, the solutions must integrate this heterogeneity and propose models capable of overcoming it.
Data, the heart of the reactor
To train the algorithms as well as possible, startups must have access to a sufficient mass of data, which represents sufficiently varied scenarios. “There are many examples where the models work with one population and not another, illustrates Philippe-Jean Bousquet. The model may be good, but the data skewed if they only considered Asian or Caucasian people, who do not have the same phenotypes, variants, exposure factors or lifestyle habits. »
Only a large amount of data from various sources can guarantee that the algorithms will not be biased. On this point, the INCa expert wants to be optimistic: “We have a powerful sandbox, thanks to the State's desire to open up health data and the presence of supercomputers in France”. For example, Primaa's algorithms have been fed by the study of "more than 100,000 slides, each allowing to obtain up to 1 million patches generating 30,000 images", specifies Fanny Sockeel, i.e. several dozen billions of data.
There remains a major challenge for companies entering this niche. "Accessing the data is not accessing the answer", warns Philippe-Jean Bousquet. The prize will not go to whoever has the largest database, but to whoever knows how to best exploit it. “Professionals who understand assumptions need those who know how to manipulate data and models, but also those who know how to industrialize this process. »
No margin of error
Unlike other AI applications, the margin of error is… zero. “When Google chooses targeted advertising on the Internet, if the success rate is 90%, the impact is almost zero for the user. If we give anti-cancer treatment to a healthy person, the impact is far from nil,” summarizes the INCa expert. Conversely, what recourse would a patient have against an AI that would miss cancer?
For this, the algorithms must be trained according to very specific patterns. “If the idea is to answer all the questions at once, the failure rate is high,” recalls Philippe-Jean Bousquet. Thus, in the case of Primaa, the company has calibrated its algorithms so that the certainty is 100% when they identify a case as negative. Conversely, this rate drops to 98% only for positive cases. “This means that we are not missing any cancer, but that we have detected a few false positives”, explains Fanny Sockeel. Since its CE marking, the startup has been working hard to further improve these statistics.
This is only possible if the AI does not become a black box incomprehensible to those who built it, like what is happening at Facebook. "You have to understand the parameters used to make them more effective, whether they are plausible or not, if it's a pattern that we know and that is valid", list Philippe-Jean Bousquet. No question of piling up the factors to be taken into account to create monsters that no one can tame anymore.
Finally, "it is not because we have the best algorithms that we have the best product", still alerts Fanny Sockeel. The last pitfall is not the least: "You have to understand what the doctor wants, present the results so that they are most useful in a simple and fast way". This requires significant ergonomic work on the applications developed, so that the user experience is optimal for the solution to be efficient.
Optimizing the choice of treatment
As in many areas, personalization also comes into play, in particular to adapt treatments to the characteristics of each patient and thus limit side effects. The Lxrepair or Resilience startups have positioned themselves in this segment. While hundreds of treatments are available on the market, "oncologists cannot know them all, but for each patient it is necessary to find the one that suits him best, the one that he will tolerate best", explained recently to La Tribune Céline Lazorthes, co-founder of Resilience. The startup recently acquired Betterise, which establishes “the profile of each patient according to a large number of criteria including medical history, treatment and how the patient reacts to it”. So much information which then makes it possible to determine the best possible treatment.
To those who fear seeing algorithms replace doctors, Philippe-Jean Bousquet reminds us that “in law, decision-making in medicine is necessarily human”. This implies a dialogue between patient and healthcare professional, and does not open the door to a decision entirely made by a machine. “Doctors are working with AI, which can bring something new. AI increases the eye of the doctor, but for very complex things, it is still necessary, observes Fanny Sockeel. On the other hand, the profession of the doctor will transform in contact with AI, he will become in a way an algorithm trainer. »
A human presence far from being confined to the role of guide, because medicine is not just a matter of statistics... "In terms of ethics, decisions taken by AI would raise many questions: if AI is wrong, who is responsible? What would be the responsibility of AI? “Algorithm certification and labeling procedures are under study, but which obviously do not settle these questions on the merits.
Better prevent relapses
This is the last brick that would make it possible to contain cancers: to better predict relapses and, therefore, to prevent them. This is what the Franco-American startup Owkin, a specialist in artificial intelligence in health, is working on – which recently signed a partnership with the pharmaceutical group Sanofi. In particular, it has partnered with the Fédération francophone de cancérologie digestive, in order to study the case of patients with digestive cancers and to better understand relapses in order to identify similarities. “Current techniques make it possible to work predictively, to identify factors that the human eye has never been able to identify and that only AI can detect. That's all the magic of AI,” enthuses Fanny Sockeel.
The cutting-edge technologies used and the diversity of the solutions developed leave you wondering: several startups, like Owkin and Resilience, even have a holistic approach, developing solutions capable of “finding” cancers from diagnosis to relapse. The rise of connected objects increases the possibilities of continuous data collection and, therefore, patient monitoring. This is why the founder of Primaa wants to be resolutely positive: “Today, it is possible to reduce cancer thanks to technology. »
Find the other articles of the file dedicated to cancer:
- What BioTechs are preparing to replace chemotherapy
- Transgene's vaccine trains the system to react to cancer cells
- Employers, here's how to approach the subject of cancer in the workplace
- 15 solutions to improve the daily lives of people with cancer
Related Articles