Hughes receives NSF CAREER Award

Michael Hughes, Ann W. Lambertus and Peter Lambertus Assistant Professor in the Department of Computer Science, recently received a National Science Foundation Faculty Early Career Development (NSF CAREER) Award for his work in machine learning and predictive modeling. The NSF CAREER Program supports exceptional early-career faculty who have the potential to drive innovative advances in research while serving as academic role models and leaders.

In this NSF-funded project, Hughes will take a multi-pronged approach to improve the training and outcomes of predictive models relating to health. In machine learning, models are trained via software that solves an optimization problem, searching over many possible model parameters to find the values that best fit a provided training data set. Today, the particular definition of what it means for a model to “fit” data well is often chosen among limited options that can be solved efficiently. However, stakeholders often have an alternative idea of how they would like to measure model fitness, but cannot easily train models directly with this definition of fitness in mind. In this project, Hughes and his research team plan to develop ‘decision-aware’ methods that can directly train models for a wider set of fitness functions. The work will entail both fundamental advances to how machine learning models are trained as well as translational efforts towards several motivating health applications.

One target application of this project is to predict where opioid overdose events will occur in the near future. Data-driven modeling could help policy makers and healthcare organizations identify specific neighborhoods in need of intervention and prioritize support for these affected communities. Hughes’ proposed work will train decision-aware probabilistic forecasting models that are specifically good at identifying the top few neighborhoods in greatest need of intervention. This work will build upon Hughes’ ongoing collaborations with Professor Thomas Stopka and Professor Shikhar Shrestha, both of the Tufts Department of Public Health and Community Medicine.

Another target application is to identify structural heart disease in an individual patient given an ultrasound imaging scan of their heart. Automating this preliminary diagnostic task could improve reliability and hopefully reach more patients compared to existing clinical practice. Hughes’ proposed work will train decision-aware probabilistic classification models that can meet specifically desired false discovery rates and false positive rates to ensure clinical utility. Because acquiring training data with associated expert labels of disease status is difficult, a key innovation will be to develop methods that can train on both the available labeled data and a much larger unlabeled dataset of only images without any labels. This work will build upon Hughes’ long-standing research collaboration with Dr. Benjamin Wessler, a cardiologist at Tufts Medical Center and assistant professor in Tufts’ Clinical and Translational Science Program in the Tufts University Graduate School of Biomedical Sciences.

Hughes earned his MS and PhD in computer science from Brown University. He joined Tufts in 2018. His research interests include Bayesian hierarchical models for documents, sequences, networks, and images; optimization algorithms and variational methods for approximate inference; and learning from labeled datasets of limited size. Ultimately, he hopes to develop predictive and explanatory models that find useful structure in large, messy datasets and help people make decisions in the face of uncertainty.

Learn more about Ann W. and Peter Lambertus Assistant Professor Michael Hughes.

The content of this article is solely the responsibility of the authors and does not necessarily represent the official views of the National Science Foundation. Research reported in this article was supported by the National Science Foundation under the following award number: 2338962