Charles Onu @Mila & Mcgill University: Applying Machine Learning in Healthcare for Social Innovation
Updated: Mar 4
Robin.ly Exclusive Interview at NeurIPS 2019
Charles Onu is a Ph.D. student and Machine Learning Researcher at Mila (the Quebec Artificial Intelligence Institute) and the Reasoning and Learning Lab (RL Lab) at McGill University. He is also a Co-Founder and AI Research Lead of Ubenwa Health, a social enterprise using AI to save newborn lives through inexpensive diagnostic technologies based on babies’ cries. In this interview, he shared his interest in both machine learning and healthcare, his motivation behind pursuing social innovation by founding Ubenwa, as well as his vision on the future direction of AI.
Charles received his master’s degree of Computer Science, specifically in machine learning at McGill University. His recent research interest includes representation learning, multi-modal fusion, few-shot learning and transfer learning at the intersection of artificial intelligence and medicine. He is also a Jeanne Sauvé Fellow and an Associate Fellow of the Royal Commonwealth Society.
Robin.ly is a content platform dedicated to helping engineers and researchers develop leadership, entrepreneurship, and AI insights to scale their impacts in the new tech era.
Subscribe to our newsletter to stay updated for more NeurIPS interviews and inspiring AI talks:
Full Interview Transcripts
1. Interest in AI Health Care
Margaret Laffan: I'm here at NeurIPS 2019 with Charles Onu. Charles, you are a social innovator and machine learning researcher. Can you tell us a bit around your interest in health care and what has motivated you to be involved in this area?
Thank you. My interest in healthcare is quite a personal one. I grew up in Nigeria. When I was growing up, it was very easy to see if you weren't above the middle class, you couldn't really afford standard and basic health care. The public healthcare system was a mess and is still a mess. Only in top level private hospitals, you could get the care you needed.
I always wanted to participate in that and become a doctor when growing up. But I realized that to become a doctor, you need to keep up with mathematics to study medicine, and I wanted to do math. So I ended up going to study engineering.
For faith, I didn't plan it, but faith just put me back into work in healthcare and engineering and machine learning. I feel very fortunate to be connected to my two loves, mathematics and healthcare, and to try to make impact with both of them.
Margaret Laffan: You're currently a machine learning researcher at Mila, the Quebec Artificial Intelligence Institute. What does your research work there involve?
My research at Mila has been broadly around the analysis of physiological signals coming from the human body, such as heart rate, respiratory activity, brain signals, and then using these to help doctors with making better decisions in the intensive care unit, and improving diagnostic systems as well.
Margaret Laffan: Can you share some more about the current research that you're doing? Is there anything that featured here at NeurIPS this year?
I do not have a paper to present this year. But last year, we presented our work in analyzing heart rate and respiratory rate signals of infants in the ICU. These infants are born premature, and they need mechanical ventilation to support spontaneous breathing. What the doctors have to decide for these babies is, when do they remove the tube that's put in their throat to support breathing, because the longer it stays in, the increasing chance of lung disease there is for the baby. If you remove it too early, it will result in mortality, the baby will die, because it might not be ready to sustain its own spontaneous breathing. So the idea was, how can we use machine learning to analyze the cardiorespiratory behavior of babies to determine when is the optimal time to win the baby from this ventilation system? This was a collaboration between Mila and the Montreal Children's Hospital.
2. Ubenwa: Use AI to Save Newborn Babies
Margaret Laffan: Can you share something about Ubenwa, the AI startup that you co-founded?
The goal of Ubenwa is to develop low-cost medical devices for resource limited settings. Our first product is a mobile app that will analyze a baby’s cries to predict signs of brain injury. This was something that came out from the work I did in Nigeria, when I was working with NGOs and realizing that birth asphyxia is one of the largest causes of mortality of babies in Nigeria and many developing countries. The reason for most of the mortality is simply because the equipment for the diagnosis is not available for the experts to use.
There’re two series of events. On one hand, I had worked previously with baby’s cries to detect emotions of the baby - happy, sad, angry or hungry. With that experience, I came across clinical research that showed that the same region of the brain coordinates speech and breathing. And because of this, we were wondering if we can use baby’s cry to diagnose asphyxia, which is one of the top three causes of mortality today. What that means is, we could deploy it into cheap devices like our phones, wearables and baby monitors, and this would replace $15,000 - $20,000 medical equipments that are simply not going to be affordable in many of the clinics and hospitals we are looking at.
So at Ubenwa, we’ve developed this app, we've collected data and used machine learning to show that this is possible. And we have a proof of concept now that we are testing in hospitals both in Nigeria and in Montreal. We have about six centers, two hospital and four clinics. The doctors are working with us now to evaluate it in the field, but also to collect more data to improve the robustness and accuracy of the system.
Margaret Laffan: When it comes to machine learning, how does that factor to what you're building out?
Charles Onu: The machine learning component is in analyzing the baby's cries, the audio signals of the baby, because the patterns that are symbols of normal babies and babies that are at risk of brain injury are quite different. But these are not things that the human ear can hear. So if you listen to two babies crying, you don't know whose crying is normal, and who is suffering from low oxygen supply to the brain.
This is where machine learning comes in. We can look into the spectrogram of the audio signals and find out the intricate details that differentiate two groups of babies. We then developed a model based on this that we used as our decision making tool. Based on past experience, we can say that this baby shows very high signs of abnormality and it's like at risk of brain injury.
3. NeurIPS 2019 and Future Trends
Margaret Laffan: At NeurIPS this year, what research is exciting you most?
Charles Onu: I've seen so many posters and talks. Many have been on the theoretical findings, which are beautiful in their own way. But what I really found fascinating is the work around vision systems, especially augmented reality systems. Snapchat had a demo, where you can go in front of it and it gives you a mask or something like that, I just found it funny and interesting. I would like to extend it to which we've pushed what is real, because this thing looks very tangible, but just a figment of the computer's imagination.
Margaret Laffan: When you think about next year in 2020, where do you see major breakthroughs for AI?
There are many interesting pathways now. One very interesting thing, which I'm also going to start looking into myself in my research, is this idea of not just treating data that we use to train machine learning as independent points, but as connected points in a cloud, and being more specific, considering data as a graph.
For instance, when you look at individuals in a social network or patients in the hospital system, they're connected in many ways, not just as independent points. The patients who have metastatic cancer are connected to some other patients by either the kind of cancer they have, the way they got the cancer, the age group, or the kind of social problems they encountered. This is one work I'm doing in cancer treatment for patients as well. Treating data more cohesively as graphs is an interesting direction of the community as a whole. And I want to try and bring this into the work I'm doing with physiological signals of patients and try to see how we can use that to build a more robust model of the patients to find out what treatments are most effective for them.
Margaret Laffan: Is that what you're most excited about for next year?
There's a bunch of things and the community is so huge. We get 13,000 people here at NeurIPS this year, at least half of which are researchers working on thousands of research directions that are all equally exciting. That's what interests me.
Margaret Laffan: Charles, thank you very much for joining us this morning.