Daphne Koller leads Insitro, a company applying machine learning to pharmaceutical drug development. She is also a co-founder of Coursera (one of the world’s largest online education platforms with 30 million + users), a MacArthur Fellowship recipient, and Computer Science Professor at Stanford University.
Recently, she discussed machine learning in the biomedical industry as well as what she thinks are vital characteristics to achieve individual success with Iridescent CEO & Founder Tara Chklovski.
Tara Chklovski (TC): Can you share two of the most interesting problems that you’ve tried to tackle in your career and what you’re working on now?
Daphne Koller (DK): My work cuts across a variety of fields including biomedical, education, and problems like computer vision. There are two specific biomedical projects I’m very proud of. The first one involved working with a neonatologist, a doctor for premature babies, to predict survival rates for neonates. Neonates are premature, teeny babies – they’re about the size of your hand and weigh less than 200 grams. They often struggle to survive. The doctor and I worked together to predict the risk of death for these babies. Our hope was to help doctors identify and help the babies at highest risk survive. In our study we didn’t rely on the traditional tests for babies’ health (which we found don’t work well for babies born that early), and we didn’t want to rely on invasive measures like sticking these tiny babies with needles! So we used the data from the bedside monitor. The bedside monitor measures things like the baby’s heart rate, respiratory rate, and their oxygen saturation. It turns out that there are markers in that non-invasive data that can predict survival rates, and that you don’t need that much data to make useful recommendations. We were able to make predictions based on data collected from the first few days of the baby’s life.
Another project I did with a PhD student of mine who was an MD/PhD pathologist. We were looking at images of tumors from breast cancer patients and trying to predict five-year survival rates to help organize patients by risk. We took a data-driven approach. Pathologists have been looking at these sort of images for a century. For our project, we included more features in our dataset than the standard features pathologists consider. We put in everything that we could think of, including hundreds of features that no one has ever looked at before. As a result, our predictions were better than pathologists’, and the features that were being used by our predictor were quite different from the ones that pathologists had been looking at. Specifically, they had been looking at features of the tumor cells, and it turned out that the features of the cells that surround the tumor were actually more predictive! Today, this is well recognized as a critical factor in cancer survival, and is called the tumor microenvironment. This microenvironment where cancer lives, and the immune system are factors that are absolutely critical in cancer patient survival. Our paper was one of the earliest pieces of evidence supporting that.