Methods & Modeling

About this day

With success defined on Day 3, today you design the model that reaches it. The morning lecture treats method selection as a reasoned choice: health-specific inductive biases, task-specific versus foundation models, when to fine-tune versus prompt versus train from scratch, representation learning, and how data structure constrains method. The afternoon lecture covers making a built model trustworthy without breaking it. The mode is build — you commit to an approach and defend it.

Schedule

Time	Block
9:00–9:15	Morning handoff
9:15–10:30	AM lecture (McDermott)
10:30–11:00	Coffee
11:00–12:30	Small Group 1
12:30–1:30	Lunch
1:30–2:45	PM lecture (Gerych)
2:45–4:15	Workshop — build your project’s methods notebook
4:15–5:45	Small Group 2 (closing)
5:45–6:00	Daily wrap — recap & look ahead
6:00	Dinner

Morning lecture outline

Methods & modeling: every choice for a reason. A working outline of the lecture’s content.

• The thesis. Each modeling choice should be dictated by an expected property of your problem and data — and the methodological work is to justify, test, and show those choices. “I want to use a transformer” is not a reason.
• What the data actually is. A patient is a longitudinal stream of timestamped, typed events (MEDS-like) — a flat feature table is already a lossy modeling choice.
• Model vs. algorithm. The loss and optimizer live in training; the model is only what predict runs — and whatever predict needs (a k-NN’s training set, a graph’s neighborhood) has to ship inside it.
• Three coupled choices. Data representation, model class, and the loss/objective — none independent; each constrains the others.
• Representation. A good representation exposes the signal (spectrograms for ECG/EEG; eGFR over raw creatinine). Pitfall: mean-imputing a missing lab throws away informative missingness.
• The label is a choice too. Predicting “needs a ferritin test” fails because the label only exists where a clinician already suspected it — selection bias / MNAR.
• Model class. Match the inductive bias: gradient-boosted trees still beat deep nets on tabular clinical data; CNNs win on images. Don’t pick by fashion.
• Loss / objective. Encode the invariance you actually believe (shape-not- appearance for imaging; a pre-training graph for relational data) — a fashionable objective with no matching property just ties plain ERM.
• Epistemic humility. For most EHR tasks we can’t yet say which choices matter — “it depends,” and we can’t say on what. That gap is the opportunity.
• Foundation models for health. MIMIC wasn’t out of data, just out of ideas; event-stream foundation models (ETHOS, CoMET) scale like LLMs.
• Think stupider, then test. Shrink the problem, synthesize data with and without the property, break the model on purpose — controlled experiments you can re-run on every change. (This feeds the afternoon synthetic-experiment notebook.)

What you’ll be able to do

• Justify a method choice against your data and your Day 3 metrics.
• Reason about foundation versus task-specific models for your problem.
• Identify your main methodological risk and a credible alternative.

Small group & workshop

In Small Group 1, the group works through the Day 4 Workbook prompt — from each project’s baselines to the structure it can exploit, and the fastest experiment to test it. See how the small groups run.

The workshop then has you build a methods notebook for your project — a baseline, stronger models, a representation / foundation-model approach, and a comparison harness that reuses your Day 3 evaluation moves. You start from a blank notebook (no template); a prompt-a-thon is the alternative track for LLM-centric projects.

The closing small group ends with each person noting how their method thinking moved over the day.

Project Workbook — Part 4: Methods & modeling

The naive and existing methods that form your baselines and the structure they leverage; where they still fall short on your Day 3 metrics; the structure of your data or problem you can exploit instead and why it should work (across pre-processing, model class, and training objective); the fastest experiment that would show your idea working or failing; and the key experimental questions for the project.

Additional resources

• Slide decks: Morning — McDermott (PDF) · Afternoon — Gerych.
• Project Workbook template — Part 4.
• The MEDS ecosystem documentation, as an implementation reference.
• Current EHR foundation models worth knowing for the foundation-vs-task-specific decision: ETHOS, CEHR-GPT, and COMET, plus the MEDS-DEV benchmark (names to look up — verify the latest before relying on them).

Suggested readings

1. Grinsztajn, Oyallon & Varoquaux (2022), Why do tree-based models still outperform deep learning on tabular data? NeurIPS.
2. McDermott et al. (2024), Using machine learning to develop smart reflex testing protocols. JAMIA.
3. Renc et al. (2024), Zero-shot health trajectory prediction using transformer (ETHOS). npj Digital Medicine.
4. Dey et al. (2025), Learning General-purpose Biomedical Volume Representations using Randomized Synthesis. ICLR.
5. McDermott (2025), The (lack of?) Science of Machine Learning for Healthcare. ML4H (PMLR).
6. Wornow et al. (2023), The shaky foundations of large language models and foundation models for electronic health records. npj Digital Medicine.
7. Steinberg et al. (2021), Language models are an effective representation learning technique for EHR data. Journal of Biomedical Informatics.
8. McDermott, Yap, Szolovits & Zitnik (2023), Structure-inducing pre-training. Nature Machine Intelligence.
9. Assran et al. (2023), Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture (I-JEPA, with LeCun). CVPR.
10. Diamant et al. (2022), Patient Contrastive Learning: a performant, expressive, and practical approach to ECG modeling. PLOS Computational Biology.
11. Ronneberger, Fischer & Brox (2015), U-Net: Convolutional Networks for Biomedical Image Segmentation. MICCAI.