Healthcare spending is nearing $3 trillion per year, but in spite of this expenditure, the US is outpaced by most developed countries with regard to outcomes. Until recently, one of the key bottlenecks for research in care delivery was the lack of data to analyze the health system’s workings. But today, post the HITECH in 2009, much of an individual’s health data is stored electronically. This opens up a wealth of opportunities for computational scientists.
In this talk, I will develop and solve two problems. The first is the challenge of accurate prognoses in chronic diseases where individuals show a great deal of variability. Here, access to accurate prediction tools can help tailor treatment decisions. We develop a probabilistic model that exploits the concept of subtypes to individualize predictions of disease trajectories. These subtypes identify groups of individuals that share a similar disease course. These subtypes are learned automatically from data. On a new individual, our model incorporates static and time-varying markers to dynamically update predictions of subtype membership and provide individualized predictions of disease trajectory.
The second challenge pertains to developing cost-sensitive predictive models. In healthcare, measurement costs share complex structure that existing cost-sensitive approaches do not tackle. We develop a new framework for defining structured regularizers that are suitable for problems with complex cost structures. Our approach is based on representing the problem costs as a multi-layer boolean circuit from which we can define our regularizer in a natural way in the spirit of group penalty functions. We show that, by incorporating ones knowledge of the underlying cost structure into the design of the regularizer, one may obtain models that are in harmony with the underlying cost structure, and as a consequence, achieve higher predictive accuracy for a given level of cost.
We've tailored Netflix recommendations and Amazon shopping carts but we're only at the early stages of knowing how to use data to tailor health decisions.
Suchi Saria is an Assistant Professor at Johns Hopkins University with appointments in Computer Science, Applied Math & Statistics and Health Policy. She received her PhD at Stanford with Professor Daphne Koller. Her research interests are in machine learning and statistical inference techniques geared towards leveraging electronic health data. Her work on risk prediction in infants was published as a cover article in Science Translational Medicine (Science/AAAS Press), and has been licensed by Nihon Kohden, the largest monitoring company in Japan. Her work has been recognized by awards including, a Best Student Paper by the Association for Uncertainty in AI, a Best Student paper finalist by the American Medical Informatics Association, a Google Research award, the Rambus Fellowship, the Microsoft scholarship and the National Science Foundation Computing Innovation Fellowship.