Today I present to you three research directions that apply the latest achievements in artificial intelligence (mostly deep neural networks) to biomedical applications. Perhaps, this is the research that will not only change but also significantly extend our lives. I am grateful to my old friend, co-author, and graduate student Arthur Kadurin who has suggested some of these projects.
Translated from Russian by Andrey V. Polyakov. Original article here.
Polar, Beyersdorf AG, and Others: Smart Clothes
We begin with a series of projects that are unlikely to turn the world over but will certainly produce, pardon the pun, cosmetic changes in everyday life already in the nearest future. These projects deal with AI applications for the so-called “Internet of Things” (IoT), specifically applications that are very “close to the body”.
Various types of fitness trackers, special bracelets that collect information about heartbeat, steps, and so forth, have long ago entered our life. The main trend at the sportswear companies now is to build different sensors directly into the clothes. That way, you can collect more information and measure it more precisely. Sensors suitable for “smart clothes” were invented in 2016, and already in 2017 Polar has presented Polar Team Pro Shirt, a shirt that collects lots of information during exercises. The plot will no doubt thicken even further when sports medicine supported by artificial intelligence will learn to use all this information properly; I expect a revolution in sports that Moneyball could never dream of.
And it is already beginning. Recently, on November 24–26, the second SkinHack hackathon dedicated to applying machine learning models to analyzing data coming from such sportswear took place in Moscow. The first SkinHack held last year was dedicated to “smart cosmetics”; the participants tried to predict the age of a person by the skin structure on photographs looking for wrinkles. Both smart cosmetics and smart clothing are areas of active interest for Beiersdorf AG (commonly known as the producer of the Nivea brand), so one can hope that the commercial launch of these technologies will be not long in coming. In Russia, SkinHack was supported by Youth Laboratories, a company affiliated with the central characters of our next part…
Insilico: Automatic Discovery of New Drugs
Insilico Medicine is a company well known in the biomedical world. Its primary mission is to fight aging, and I personally wish Insilico success in this effort: one does not look forward to growing old. However, in this article I would like to emphasize another, albeit related, project of the company: drug discovery based on artificial intelligence models.
A medicinal drug is a chemical compound that can link with other substances in our body (usually proteins) and have the desired effect on them, e.g., suppress a protein or start producing another in larger quantities. To find a new drug, you need to choose from a huge number of possible chemical compounds exactly the ones that will have the desired effect.
It is clear that at this point it is impossible to fully automate the search for new drugs: clinical trials are needed, and one usually starts testing on mice, then on humans… in general, the process of bringing a new medicinal drug to the market usually takes years. However, one can try to help doctors by reducing the search space. Insilico develops machine learning models that try not only to predict the properties of a molecule but also to generate candidate molecules with desired properties, thereby helping to choose the most promising candidates for further laboratory and clinical studies.
This search space reduction is done with a very interesting class of deep learning models: generative adversarial networks (GAN). Such networks combine two components: a generator trying to generate new objects — for example, new molecules with desired properties — and a discriminator trying to distinguish generated results from real data points. Learning to deceive the discriminator, the generator begins to generate objects indistinguishable from the real ones… that is, hopefully, actually real ones in this case. The last Insilico model, called druGAN (drug + GAN), attempts to generate, among others, molecules useful for oncological needs.
MonBaby: Keeping Track of the Baby
Finally, I would like to end with a project that Neuromation plans to participate in. Small children, especially babies, cannot always call for help themselves and require special care and attention. This attention is sometimes required even in situations where mom and dad seem to be able to relax: for example, a sleeping baby may hurt a leg by turning to an uncomfortale pose . And then there is the notorious SIDS (sudden baby death syndrome), whose risk has been linked with the pose of a sleeping infant: did you know that the risk of SIDS increases several times if a baby sleeps on the stomach?
The MonBaby smart infant tracking system is a small “button” that snaps onto clothing and monitors the baby’s breathing and turning around while asleep. Currently, the system is based on machine learning for time series analysis: data from baby movements is used to recognize breathing cycles and sleeping body position (on the stomach or on the back).
We plan to complement this system with smart cameras able to track the infant’s movements and everything that happens to him or her by visual surveillance. The strong suits of our company will come in handy here: computer vision systems based on deep convolutional networks and synthetic data for their training. The fact is that in this case it is practically impossible to collect a sufficiently large real data set for training the system: it would take not only real video recordings of tens of thousands of babies, but video recordings with all possible critical situations. Thankfully, modern ethics, both medical and human, would never allow us to generate such datasets in real life. Therefore, we plan to create “virtual babies”, 3D models that will allow us to simulate the necessary critical situations and generate synthetic videos for training.
We have briefly examined three directions in different branches of biomedicine — sports medicine and cosmetics, creating medicines and baby care — each of which is actively using the latest achievements of artificial intelligence. Of course, these are just examples: AI is now used in hundreds of diverse biomedical projects (which we may touch upon in later articles). Hopefully, however, with these illustrations I have managed to show how AI research is working on helping people live longer, better, and healthier.
Sergey Nikolenko,
Chief Research Officer, Neuromation
Leave a Reply