Instructor Interview
Below, Leo Celi describes various aspects of how he and his colleagues taught HST.953 Clinical Data Learning, Visualization, and Deployments in the fall semester of 2024.
OCW: Please describe the sharing of responsibilities among the members of the teaching team, and how you coordinated your efforts.
Leo Celi: We literally had an army who helped with the workshops and the projects. These were visiting students, scientists, and postdocs at our lab. They represented a gamut of backgrounds from data science, medicine, and social science, across generations from Gen Z to Gen X. Teaching and learning from the course was part of their MIT experience. They helped design and execute the workshops and plug into teams with the students taking the course.
OCW: What did you hope students would take away from their experiences in the course?
Leo Celi: The most important lesson we wanted the students (and the teaching team!) to take away from the course is that the data is not an objective representation of biological patterns of health and disease. The way the data was collected, with what devices, the choice of what to collect as proxies of what we are measuring, the way healthcare is delivered—who makes it to the hospitals and who doesn’t—are all shaped by structural issues that are far too complex for an individual or even a large group of researchers. The flagship offering of our lab is the MIMIC database with more than 100,000 users from around the world. More than 15 years after its release, we are still discovering “data irregularities,” e.g., patterns of missingness, variations in documentation habits outside of missingness (e.g., likelihood of rounding off measurements), because tens of thousands of pairs of eyes are scrutinizing the data. Algorithmic bias is a product of a multitude of upstream issues even before data collection (e.g., epistemology of medical devices), and the only way to mitigate that is to bring together people who can look at the data from as many lenses as possible. Sadly, the 2024 instance of HST.953 is likely the last time we will offer the course, as it has been supplanted by a new IMES course on AI in healthcare.
OCW: To what extent did you follow a pre-established plan for the semester? How much flexibility did you allow yourself in implementing the syllabus?
Leo Celi: We had last minute “substitutions” during the semester. Given how quickly the field has been moving (and will continue to do so), we have learned to be more agile when it comes to what we offer. We constantly come up with new exercises and activities to adapt to new challenges and topics in the field.
OCW: As the syllabus notes, the expectation was that by the end of the course the final project would be developed enough to submit to a peer-reviewed journal. How have the students’ projects fared in the wider world? Have many of them seen publication in such journals?
Leo Celi: Most of the projects eventually get published, but with a huge variation in who stays through the finish line among those who take the course. This is why notebooks and repos are the most important deliverables that we require from the students.
OCW: What would you like to share about teaching HST.953 that we haven’t yet addressed?
Leo Celi: We realized early on that what we need to teach is not specific concepts or specific skills. We need to teach students how to learn, which the vast majority of courses don’t do well. Students expect to be taught what they need to learn, but we, the teachers, actually don’t know how to learn, especially in this age when knowledge is everywhere. What we need to teach is the agency to situate the knowledge that is delivered by AI, because not knowing how that knowledge came about makes the application of that knowledge “risky.” Rather than using the knowledge, the knowledge (or more accurately, the creators of that knowledge) may be using us. The design of our courses, our events, our workshops, was reflexive and reflective, continuing to evolve from year to year based on what we observed, what we and the learners experienced.
Assessment
Grade Breakdown
The students’ grades were based on the following activities:
- 15% Weekly reflections
- 50% Problem sets
- 35% Final project
Curriculum Information
Prerequisites
Students were recommended to have first taken the following courses, or to have some equivalent experience in machine learning, visualization, and human-computer interaction:
- 6.867 Machine Learning
- 6.S897/HST.956 Machine Learning for Healthcare
- 6.813/6.831 User Interface Design and Implementation
Offered
HST.953 was offered most fall semesters from 2016 through 2024.
Student Information
Enrollment
93 students
Breakdown by Academic Program
Two-thirds to three-quarters of the enrolled students were HST students from Harvard Medical School or were studying at Harvard’s T.H. Chan School of Public Health; the remainder were MIT graduate students.
Typical Student Background
About half of the students had directly relevant prior experience in the course topic.
How Student Time Was Spent
During an average week, students were expected to spend 12 hours on the course, roughly divided as follows:
In Class
Met 1 time per week for 2.5 hours per session; 13 sessions total; mandatory attendance.
Out of Class
Outside of class, students completed assigned readings, wrote weekly reflections, worked through problem sets, and collaborated in teams on the final project.
