“One could say that a man can ‘inject’ an idea into the machine, and that it will respond to a certain extent and then drop into quiescence, like a piano string struck by a hammer.” – Alan Turing, 1950.
At the time computers were first invented, Artificial intelligence (AI) was an unimaginable concept. The first computer was invented by British mathematician Charles Babbage at the end of the 19th century. The machine was simply a giant “Analytical Machine” (1) that computed simple arithmetic at his command, a 500 kilogram calculator. It wasn’t until the middle of the 20th century, a timespan of approximately 50 years later, that computers advanced from metal abacuses to machines capable of learning patterns, decoding messages, and solving complex questions. Computer pioneer Alan Turing, who created the "bombe" machine that broke the seemingly unbreakable Nazi Enigma code, was obsessed with the question, “Can machines think?” Integrating philosophical questioning to his work, he firmly believed that no matter how we define the human nature of cognition, by the 21st century, machines would have the capability to “imitate” humans. (2)
The growing field of AI can be both exciting and unsettling. As with anything novel, it takes years of exploration for us to find the boundaries and limitations. AI is now widely implemented in many aspects of our daily lives—TikTok video recommendation algorithms, automatic Facebook bots, and facial ID to unlock smartphones, to list a few. As ubiquitous and invisible as AI has become, it is not surprising that AI has been used by some to invade our privacy and even attempt to manipulate our behaviour. No wonder that some have developed a deep-rooted distrust in AI. This technology is unfamiliar to most people and is advancing into everyday life at a brisk pace.
That said, we tend to forget that AI is not a pre-existing machine but the spawn of a human idea. All AIs start with one line of code and one data point. How AI is used to impact our lives depends largely on the motives of its creators and implementers. While the purpose of AI is misconstrued by some, there is no doubt that when AIs are designed and implemented to help us, their benefits far outweigh their shortcomings. As a scientist working closely with AI and therapeutic discovery, I want to highlight the recent advancement of AI as applied to medical sciences.
AI in Genomic Studies
Ever since the human genome was completely mapped in 2003, scientists have been leveraging new computational power combined with the large deposit of genomic data to understand complex human diseases. Among them is Dr. Chloé-Agathe Azencott (3), a leading scientist who focuses on machine learning in drug discovery. Scientists like Dr. Azencott use machine learning algorithms on our genome to predict disease “phenotypes”which are physiological traits that can be observed or measured. The algorithms iteratively improve scientists’ predictions until they find the real phenotype and reach a minimum error. Through this process of learning, the algorithms can help us find out what genes are involved in a disease or in the response to a treatment. The algorithms can piece together samples from multiple people and consider many phenotypes all at the same time—a powerful ability that human minds simply do not have.
In my conversation with Dr. Azencott, I raised the question of the applicability of genomic studies to all people—I suspected that the social inequity in healthcare infrastructure and technology across the world would limit our understanding of gene-disease interaction to only those whose genomes are sequenced. She agreed:
“Most of what we know (except in breast and ovarian cancer) pertains to white men from Western Europe or Northern America…The latest estimation I've heard is that Europeans make up 80% of all [genomic] data but about 16% of the world's population.(4) This means that we don't know whether these findings apply more widely. Actually, we even know that they don't. (5) (6)”
Even the most powerful AI cannot magically resolve these limitations. “In my group, we've been working on techniques based on what is called multitask machine learning to combine data sets from different populations,” says Dr. Azencott. “These approaches let us find both loci (genomic locations) that are relevant for all populations and population-specific loci. They can help analyse data where we have fewer samples from underrepresented populations…[but] they can't invent things about populations for which we have no samples at all.”
Much work is still needed to refine these algorithms. We have found thousands of human genetic variants associated with disease, but often these predictions and hypotheses only explain a fraction of the complex traits we see. For example, most current algorithms consider each change within the genomic code to be completely independent, thus neglecting the possibility that these changes interact with each other. “That's why we need algorithms such as those I develop because we consider multiple loci jointly,” Says Dr. Azencott. By continuously refining AI methods to be more complex and inclusive, we hope to interpret genomic data with more confidence. The outlook of AI in genomic studies remains positive.
AI in Medical Imaging
Medical image analysis, an often complex and time-consuming aspect of patient care, is now easier with the help of many emerging AI software. Among them is DIVA, a software created here in France in collaboration with Institut Pasteur, Institut Curie, and PSL. DIVA (Data Integration and Visualization in Augmented and Virtual Environments) can visualize and analyze any 3D image stack using Virtual Reality (VR). Like dropping the yellow avatar in Google Street View to inspect your surroundings, DIVA places the user within the 3D images for point-of-view exploration and interaction. Inside the images, the user can freely change the colors and transparencies of certain features to track them and thus understand them in a more comprehensive manner.
Avatar Medical, a startup co-founded by Pasteur scientists Mohamed El Beheiry and Jean-Baptiste Masson in 2020, has been utilizing DIVA technology to help surgery planning in a variety of medical fields including oncology and orthopedics. The software allows surgeons from various specialties to easily visualize complex medical images such as MRI and CT-scans as if the patient was right in front of them, even if the surgeons have no previous background with image analysis. The machine learning algorithm allows the user to manually highlight a feature such as a cancerous tumor or a crack in a bone. The machine will remember what that feature “looks” like and automatically classify the rest of the images in the same way. This allows doctors to follow their points of interest throughout hundreds of images that make up the 3D volume. “I can now see a complete knee in the blink of an eye.” Says Dr. Greg Sarin, an orthopedic surgeon who has levfffffferaged the power of DIVA in his surgeries. Outside of hospital settings, DIVA has also been useful in the classroom. For introductory medical students with no previous surgical experience, DIVA provides a deeper understanding of anatomy and “hands-on” training without having to interact with actual patients.
DIVA promises to be as affordable as it is powerful. With increasing availability of consumer-based VR equipment such as the Oculus Rift and Windows Mixed Reality, VR-based AI technology is as accessible as ever. “The software runs on simple laptops [, so] the cost of DIVA [for healthcare practitioners] is rather limited.” Says Dr. Masson. With the help of cutting-edge AI technology like DIVA, standardizing accurate, precise, and affordable surgical care is within reach.
The future of AI in healthcare
When AI was progressing rapidly at the turn of the first decade of the 21st century, some engineers became overenthusiastic and overconfident in its potential to replace physicians in the healthcare industry. However, it seems the more we learn about something, the more we realize how little we know. Through my conversations with Dr. Azencott and Dr. Masson, we all agreed that AI is not a magical antidote to all ills, and physicians cannot simply be replaced. While AI has proven to be a powerful analytic, in the end, its ability to make executive decisions as humans would do is still up for debate. Nonetheless, we are optimistic about the use of AI as an assistance to medical practitioners. For example, AI can help us automate minute tasks such as annotating medical images, which will provide doctors more time to focus on difficult cases. AI can also support clinical decision making. For example, machine learning can calculate the best matches between kidney donors and receivers to maximize the number of kidney transplant operations and minimize the chance of tissue rejection. It will take time to prove the efficiency of AI in a realistic healthcare infrastructure, but “it is an exciting time,” says Dr. Masson. “I am convinced that interesting results will be demonstrated at the frontier of AI and medical doctor expertise.”
1. Harris, William. (2021) “Who Invented the Computer?” HowStuffWorks Science. https://science.howstuffworks.com/innovation/inventions/who-invented-the-computer.htm. accessed 24 Jan 2022.
2. A.M. Turing. (1950) Computing Machinery and Intelligence. Mind (LIX)236. https://doi.org/10.1093/mind/LIX.236.433
3. Machine learning for therapeutic research. https://cazencott.info/
4. Genetics for all. (2019) Nat Genet 51, 579. https://doi.org/10.1038/s41588-019-0394-y
5. Martin, Alicia R et al. (2019) “Clinical use of current polygenic risk scores may exacerbate health disparities.” Nature genetics vol. 51,4: 584-591. https://doi.org/10.1038/s41588-019-0379-x
6. Carlson, Christopher S., et al. (2013) "Generalization and dilution of association results from European GWAS in populations of non-European ancestry: the PAGE study." PLoS biology 11.9. https://doi.org/10.1371/journal.pbio.1001661
This article was specialist edited by Dr. Cristophe Zimmer and copy edited by Cliff Shoals.