Machines to make sense of our health

Samuel Kaski, Professor of Computer Science:

My research focuses on machine learning algorithms, which are powerful AI tools that benefit many areas in our day-to-day lives. They power things like automatic text translation, face recognition in pictures, and voice assistants like Siri and Alexa. There’s hope that soon these powerful computer-driven predictions will be put to work in hospitals assisting clinicians in making diagnoses ― a method commonly referred to as personalised medicine. If we can make a deep learning algorithm that can spot your face in a busy crowd, can we detect when you’re getting ill and identify how to best treat you? 

Cancer samples make up one successful example. Using algorithms, we’ve been able to find out what information is relevant in datasets, and in their dependencies. These tools improve prognoses of what kinds of therapies are effective for each patient, based on a specific tissue sample. 

However, one of the main challenges we face in this task is data quantity. Successful deep learning methods current require huge datasets to train on. Before a computer can detect a face in a picture, it needs to see thousands of photos of faces and more of things without faces before it can tell them apart. For patient data, though, the datasets we have to train with are much smaller. For some rare diseases there will only be information on a handful of cases ― we need to develop new methods that can provide the great predictive power of deep learning algorithms, but with much less training data. To create algorithms that would mean a computer could monitor your own health, the dataset available shrinks even further to only one person: you! My research group is working on designing new approaches better suited to smaller data sets. 

As well as being able to handle smaller amounts of data, we need to be able to make algorithms explain what they do. When your smartphone’s keyboard suggests a word to you based on what you typed, you don’t really think about how it reached that suggestion. However, if Siri started suggesting that you need to take expensive medication with side effects, or undergo a risky operation with a long recovery time, you’ll definitely want to know why it this suggestion. I co-supervise a group that is looking to develop AI methods that can explain how they have reached their conclusions, which will greatly help integrate the strength of new artificial intelligence tools into existing healthcare infrastructure.  

These problems are not unique to health applications for artificial intelligence. The best way that we can develop AI for a full range of advanced applications is by building expert communities that can work on different aspects in close collaboration. I am the director of the Finnish Center for Artificial Intelligence (FCAI), a joint initiative by Aalto University, Helsinki University and VTT. Our slogan is that we create ‘Real AI for Real People in the Real World’ and our work on developing new AI methods for personalized medicine is an example of this. FCAI brings together engineers and scientists with expertise and experience across a number of fields to develop the next generation of AI together ― and use it to solve society’s grand challenges.