Day 1 · Mon · Jun 22 · Frame

Problems in ML4H

An interesting topic → a real, well-posed health problem with stakeholders.

AM lecture: Peter Szolovits (MIT) PM lecture: Emily Alsentzer (Stanford)

About this day

Day 1 sets the frame for the week. After a short opening block — welcome, orientation to how the week works, and a brief icebreaker — the morning lecture asks what makes machine learning for health a distinct discipline, what is shared across all health problems regardless of domain, and what AI cannot do. The afternoon lecture walks a real project from messy problem-formulation toward clinical use. The work of the day is to stop thinking about your project as a method and start thinking about it as a problem: an unmet need, with stakeholders, and a reason it is still open.

Schedule

Time	Block
8:30–9:15	Opening block — check-in, welcome & orientation (Tristan Naumann), brief icebreaker
9:15–10:30	AM lecture (Szolovits)
10:30–11:00	Coffee
11:00–12:30	Small Group 1 — problem framing
12:30–1:30	Lunch
1:30–2:45	PM lecture (Alsentzer)
2:45–4:15	Specialty Breakout #1 — problem clinic
4:15–5:45	Small Group 2 (closing) — refine your problem framing
5:45–6:00	Daily wrap — recap & look ahead
6:00	Dinner

Morning lecture outline

Problems in ML4H: what we’re actually trying to solve. A working outline of the lecture’s content.

The week’s spine. Idea → problem → data → evaluation → method → deployment. Today closes the first gap: an interesting topic becomes a well-posed problem.
Clinical and population context. What clinicians actually do (diagnose, prognose, treat, monitor) and what “health” even means — the WHO definition, why social well-being is so hard to measure, and a quick tour of population-health data.
The central claim. ML4H is a distinct field, not “ML applied to health data”: labels are endogenous to the care process and deployment is performative (it changes the data-generating process), breaking the i.i.d. and identifiability assumptions ML quietly relies on.
Fifty years of perspective. From MYCIN/INTERNIST expert systems through statistical ML and deep learning to today’s foundation models — every era promised autonomy and delivered, at best, decision support.
Six features shared across health problems. High stakes; distribution shift; label noise; causal structure; regulation; a human in the loop.
Why health breaks generic ML. IID is rarely true; the label is often a proxy; “performance” is not “benefit”; error costs are asymmetric and patient-specific.
Where AI is, and isn’t, well-suited. Strong at pattern recognition at scale, triage, and surfacing missed signal (e.g. mammography-based risk); poor where the problem is actually causal, where there is no credible label, or where the binding constraint is action, access, or trust.
Prediction, causation, decision. Three different questions projects routinely conflate — even “pure prediction” needs a written estimand. (Day 3 returns to this.)
Problem formulation is a choice. A problem is not a method; the framing — target, population, label — bakes in fairness and harm before any model exists (the cost-versus-need example). Name who is affected, who decides, who pays, and separate what is genuinely new from recurring hype.

What you’ll be able to do

Distinguish a research problem from a method.
Identify the stakeholders and the unmet need behind your project.
Locate your project in the ML4H landscape.

Small groups & specialty breakout

Day 1’s first small group opens with introductions — yourselves first, then a short pitch of the project you came with. Keep the pitch brief: if your groupmates don’t quite get it from a quick description, that is useful signal about how the problem is being communicated. The group then works through the Day 1 Workbook prompt together, reframing each project as a problem — an unmet need with stakeholders — rather than a method. See how the small groups run.

The specialty breakout then runs as a problem clinic within your data-modality track: what distinguishes a strong from a weak problem in your domain, and the recurring ways problems in it are mis-posed.

The closing small group ends with each person taking a few minutes to say how their problem framing shifted over the day.

Project Workbook — Part 1: Problem

Using the Project Workbook template: your one-line pitch, what the problem is, who the stakeholders are, why it is impactful, why it is technically interesting, why it is hard, and why it hasn’t been solved before. Resist naming a method — this part is about the problem.

Additional resources

Slide decks (PDF): Welcome & orientation · Morning lecture — Szolovits · Afternoon lecture — Alsentzer.
Project Workbook template — copy it and complete Part 1.
The CHIL and ML4H proceedings, for a sense of the field’s scope.
Your specialty-breakout track materials (posted before the camp).