In Spring 2019, I (along with Irene Chen) TA’d the MIT’s first Machine Learning for Health class (see MIT News Article). You can visit the course website here to see the schedule/resources.

Exciting things about the class:

This was a very intensive class, requiring:

I’m very proud of how everything went! Great course staff! Great students! Great invited speakers!

What am I most proud of?
The psets. I spent the largest amount of my time making the psets clear and interesting. There were some pretty stressful points as deadlines approached & IT infrastructure issues needed addressing. But I’m very happy with how the assignments came out!

I worked very hard to make sure the psets were clear, both in explanation to and in expectation of students. Every time we gave the students a real dataset, we had them look at the data and try making the prediction themself to better appreciate whether the ML problem was feasible and sensible.

I’m especially proud of PS6, which used causal inference techniques to estimate the effect that prescribing patterns have on opioid addiction: it was very important to me that we didn’t jump right into applying forumlas blindly and circling the answer at the end. By asking “naive” questions first, we primed the students to think about how the formulas refined upon simple intuitive ideas.

And then from there, it was very important to get the students to consider heterogeneity in the population (meaning that while the treatment might have a large impact for one subpopulation, other groups might be less-affected). Although I think the sensitivity analysis question could’ve come out better, I’m overall extremely happy with this pset!

What did I learn?
I remember during pset 1, there was a piazza question about whether the students needed to normalize their features before fitting a model. I replied very quickly to the question telling the students they didn’t have to worry about that. About 1-2 hours later, another student (not noticing my “official” answer) replied to the OP saying they probably should normalize features because of the feature analysis question. That student was right. But because I’d already posted an official answer & some people had already finished/submitted the assignment on that assumption, I couldn’t reverse course.

After that, I realized that even though I might feel a need to swoop in and help ASAP, I should move a little more cautiously & make sure I understand the implications of whatever decisions I make (because people will make their decisions based on those answers). That general principle has stuck with me: when people depend on you, then you need to take that very seriously.

What do I wish we’d done better at?
As a class, we had 11 guest speakers, though only 4 were women (36%). I’m very happy with how our guest lectures went, but I suspect we could’ve had just as excellent discussions with 50% (or higher) female speakers.

How well did I do?
Student feedback for me was mostly positive (though I was the lower-scoring TA on the evaluations).

The written responses said my strengths were patience, kindness, responsiveness, and interest in the material. My experimental recitations were not universally well-received, which is good feedback to get.

Student responses of me overall. Only 14 responses because only 6.S897 students were able to fill out the Course 6 evaluation, Everyone enrolled as HST.956 could not write an evaluation.

What would I do differently a second time?
Multiple students expressed interest in a pset about Computer Vision for Healthcare (e.g. Chest X-Rays). This is an interesting idea, and I’d be open to exploring it. Of course, it would have some challenges w.r.t. processing power available to all of the students.

It might also be fun if there was a pset where each student needed to interview a doctor to learn about what they do & brainstorm about how ML would or wouldn’t help with various tasks.

In addition, I still think my recitations (e.g. observational data and challenges in evaluation) could be different from the lecture material but still interesting. I was hoping to use my recitations to discuss the broader context of healthcare (including regulations, deploying systems, and health policy). I suspect that there is still a way to integrate those lessons into the supplementary (optional) recitations to bring a different perspective for ML engineers in the class. Maybe I just needed to find a better set of areas to explore and/or ways to discuss that information. I tried assigning 3 podcasts/videos listen to & having a group discussion but no one listened to them!  🤣


One thought on “TA Reflection: S19 ML for Health

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s