Overcoming Limited Data Challenges to Diagnose Heart Disease and Forecast COVID-19 Hospital Trajectories

September 23, 2021
3:00 - 4:00 pm ET
Halligan 102, Zoom
Speaker: Michael Hughes
Host: Liping Liu


I'll present recent work from my lab at Tufts on two pressing problems in healthcare -- forecasting the day-by-day status of hospitalized COVID-19 patients and diagnosing heart disease from ultrasound images. Both problems have limited available data and do not fit neatly into standard machine learning tasks. I'll highlight how our proposed solutions combined methodological innovations with careful collaboration with clinicians to ensure models are informed by domain knowledge and deliver actionable insights. Both studies resulted in publications at the Machine Learning for Healthcare (MLHC) conference in August 2021.

In the first half, I'll talk about developing a flexible, mechanistic model of how COVID-19 patients move through the various stages of care in the hospital from the general ward, to critical care, and on to mechanical ventilation. I'll show how this model can produce detailed forecasts of how many beds might be needed in the future, how it can be adapted to a region of interest (e.g. Massachusetts or California) using Approximate Bayesian Computation, and how we are working with several major pharmaceutical companies to use this model to assess the societal value of possible interventions. Relevant paper: Visani et al. MLHC 2021.

In the second half, I'll talk about our work on methods to diagnose heart valve disease from ultrasound images. Working with a cardiologist at Tufts Medical Center, we are exploring how automated initial screening can help identify patients so they can be treated as early as possible. We'll discuss our methods for learning from limited labeled data. We will also introduce our newly released deidentified dataset - the Tufts Medical Echocardiogram Dataset (TMED) -- which we hope catalyzes innovation in this space. Relevant paper: Huang, Long, Wessler and Hughes MLHC 2021.


Michael C. Hughes ("Mike") is the Ann W. Lambertus and Peter Lambertus Assistant Professor of Computer Science at Tufts University. Mike leads a team focused on statistical machine learning and its applications to healthcare, with research interests spanning learning from limited labeled data, approximate inference algorithms, and Bayesian hierarchical models for heterogeneous data (especially time series, networks, and images). You can find links to recent papers and code www.michaelchughes.com.

Please join meeting in Halligan 102 (in-person) or via Zoom.

Join Zoom Meeting: https://tufts.zoom.us/j/97183120811

Meeting ID: 971 8312 0811

Password: See colloquium email

Dial by your location: +1 646 558 8656 US (New York)

Meeting ID: 971 8312 0811

Passcode: See colloquium email