SPEAKER NOTES — Slide 1 (Title) Full ~75-min lead lecture (9:15), BUILD mode; guest Walter Gerych at 1:30; you anchor the afternoon build session. The whole deck has one core message: every modeling choice should be made for a reason dictated by your problem and data — and methodological work is about justifying, testing, and SHOWING those choices. That message lands in the afternoon: students build their own synthetic-experiment notebook. Keep it lean and concrete.

SPEAKER NOTES — Slide 2 (Who am I) ~60–90s, humble. You live in both the data-infrastructure and modeling worlds — which is why the talk takes the data-generating process seriously. Be honest that MEDS is mature but the FM work is open; that honesty motivates the "test your choices" message later. Say it aloud: you're hiring PhD students & postdocs for your lab at Columbia DBMI — invite people to come find you.

SPEAKER NOTES — Slide 3 (Event stream) Establish the data view: everything downstream is a choice about how to handle this stream. A flat feature table is already a lossy projection — i.e. a modeling choice.

SPEAKER NOTES — Slide 4 (Divider — terminology) Set up part (a): nail the terminology before decomposing the design. The algorithm/model distinction is the foundation for everything after.

SPEAKER NOTES — Slide 5 (Two APIs) — CORE Two different signatures → two different objects. train(dataset,config)→model vs predict(sample)→prediction. The crux for the next slide: predict only gets a SAMPLE, so anything else the model needs must ship inside it. Also: the loss is an argument to train — part of the algorithm, not the model.

SPEAKER NOTES — Slide 6 (What ships — setup) — CORE Pose the question. predict() only ever gets one sample, so whatever else the model needs has to travel inside it. Build the answer across the next three reveals. Terminology aside to say aloud: "I'll use 'model' loosely — a class, a fitted artifact, or the whole recipe — infer from context."

SPEAKER NOTES — Slide 7 (What ships — weights) — CORE The easy one: every model ships its weights. That's the obvious part everyone pictures — set it up so the next two reveals feel like "...and what else?"

SPEAKER NOTES — Slide 8 (What ships — + reference set) — CORE A k-NN / retrieval model also ships its reference set: the training examples ARE part of the model, not just the weights. predict() needs them in hand.

SPEAKER NOTES — Slide 9 (What ships — + graph) — CORE A population-graph model also ships its graph neighborhood — its unit of prediction is "patient-in-graph," which rules out point-of-care use on a lone patient. The API fixes where the model can run.

SPEAKER NOTES — Slide 10 (Divider — three choices) Part (b). The whole design space is three decisions — and they are NOT independent.

SPEAKER NOTES — Slide 11 (Three choices) A (representation) chooses what the data become; B (model class) chooses what functions are possible; C (loss) chooses which model you actually get. Stress the coupling: a representation gates which classes are even available, and the loss only means something given the class. These three icons recur on the example slides as a visual atlas.

SPEAKER NOTES — Slide 12 (The core message) The thesis of the whole talk, stated as the rule: every choice needs a why, and the why must come from your data and problem. Everything after is examples of that, done well and badly. "Justify, test, show" sets up the synthetic-experiment breakout.

SPEAKER NOTES — Slide 13 (Do we know enough) Right after the rule: the honest caveat. Because we lack general laws, you must test choices empirically in your own setting — which motivates the examples that follow and the synthetic-experiment breakout.

SPEAKER NOTES — Slide 14 (Representation — setup) — A Frame it: representation is the first design choice. Build the win, then the pitfall, across the next two reveals. Icon = the representation glyph (recurs across A slides).

SPEAKER NOTES — Slide 15 (Representation — win) — A Spectrogram / eGFR: intelligent transforms that expose the signal (classical).

SPEAKER NOTES — Slide 16 (Representation — pitfall) — A The pitfall: naive imputation throws away the informative-missingness signal Kohane quantified.

SPEAKER NOTES — Slide 17 (Tokenization — the raw view) — A Tokenization is the whole problem of turning this raw event stream into the sequence your model actually ingests — not just "how to encode a number." Start from the input we have, and ask how to serialize it into tokens.

SPEAKER NOTES — Slide 18 (Tokenization — the design space) — A The whole design space in one figure (Lee et al., MLHC 2026): granularity / anchoring / fusion, value & temporal encoding, vocabulary grouping — all evaluated against shared outcomes & metrics. Let it land big, then walk the axes.

SPEAKER NOTES — Slide 19 (Tokenization — axes) — A Walk the figure's axes briefly: granularity/anchoring/fusion, value & temporal encoding, vocabulary grouping — each measurably changes a model. The framing point: this is event-stream → model input, not just "encode a number." Keep it short — the figure carries the detail.

SPEAKER NOTES — Slide 20 (Model class — setup) — B Frame it: the model class is a structural claim. Build the win then the pitfall. Icon = model-class glyph.

SPEAKER NOTES — Slide 21 (Model class — win) — B Trees win on tabular; CNNs win on images — the bias matches the structure.

SPEAKER NOTES — Slide 22 (Model class — pitfall) — B The pitfall: a fashionable deep net on tabular loses. The reason is the bias, not the brand. Leads into the quick-hits tour of architectural biases.

SPEAKER NOTES — Slide 23 (Quick hits — CNN) — B CNN: weight-sharing / translation equivariance, and the Fourier (convolution = pointwise multiply) view. Reveal one architecture at a time.

SPEAKER NOTES — Slide 24 (Quick hits — Transformer) — B Transformer: all-pairs attention is a length-scaling memory with no fixed hidden-state bottleneck (contrast RNNs).

SPEAKER NOTES — Slide 25 (Quick hits — ResNet) — B ResNet: identity skips ease optimization and reframe each layer as a residual correction.

SPEAKER NOTES — Slide 26 (Quick hits — U-Net) — B U-Net: cross connections preserve high-resolution detail for segmentation. The architecture IS the inductive bias — pick it to match the data's structure.

SPEAKER NOTES — Slide 27 (Loss concept) — C Reframe: the loss encodes an inductive bias about the task and data, and the job is to justify why that bias should help — not to claim we know the truth. Two defensible examples follow (Dey et al. shape-not-appearance; SIPT structure-inducing pretraining). The pitfall: a fashionable objective that doesn't match your data may not last (examples to come). Icon = loss glyph.

SPEAKER NOTES — Slide 28 (Loss example: shape, not appearance) — C Your favorite example (Dey et al., Golland lab). The data engine + dense contrastive loss literally encode "shape, not appearance": the same anatomy across randomized renders are positives. Payoff: general-purpose 3D features with no real pre-training data, strong on registration and few-shot segmentation.

SPEAKER NOTES — Slide 29 (Loss example: SIPT) — C Your own work. SIPT frames pretraining as inducing a graph structure over samples, then optimizes for a desired structure directly — writing the relational prior into the objective. Complements the imaging example: both put a known invariance/structure in the loss rather than hoping it emerges.

SPEAKER NOTES — Slide 30 (MIMIC — setup) A benchmark we declared finished supported a whole new paradigm. State the claim first, then reveal how autoregressive EHR FMs actually work.

SPEAKER NOTES — Slide 31 (MIMIC — how AR FMs work) Show HOW autoregressive EHR FMs work: roll out K×T sampled timelines, then estimate. ETHOS/ARES are MIMIC; the scaling complement (next) runs on Epic Cosmos.

SPEAKER NOTES — Slide 32 (MIMIC — scaling) Same point, second figure: event-stream FMs improve along clean power laws (CoMET / Epic Cosmos). "Saturated" was about ideas we hadn't had yet, not the data.

SPEAKER NOTES — Slide 33 (EHR modeling timeline) What we treat as the right structural framing of EHR data has itself evolved repeatedly — disease-progression → next-visit prediction → pretrained representations → zero-shot trajectory / foundation models. The data didn't change, the framing did.

SPEAKER NOTES — Slide 34 (EveryQuery — model) Speculation: the timeline rests on one inductive bias — the temporal structure of EHR data. EveryQuery (ours) bets we're not done with it: instead of sampling K×T autoregressive rollouts, condition on a query and predict directly — one forward pass, continuous output. Show the AR-vs-EveryQuery contrast figure.

SPEAKER NOTES — Slide 35 (EveryQuery — results) Same framing, the payoff: across MIMIC-IV, NWICU, and AMC, EveryQuery matches or beats autoregressive rollouts at thousands of times the inference speed (and higher AUC) — evidence the temporal inductive bias still has room to give.

SPEAKER NOTES — Slide (Divider — validating choices) Pivot: we've decomposed the choices and seen examples, but we can't assume any of them are right. The rest — and the afternoon — is about how to TELL: shrink the problem, run synthetic experiments, and turn them into re-runnable tests.

SPEAKER NOTES — Slide 37 (Think stupider) This is the method for testing choices, and the design pattern for the afternoon notebook. Surprises in these stripped-down settings are where the science is, because everything else is controlled. Give one concrete example from your work.

SPEAKER NOTES — Slide 38 (A good synthetic experiment) The design pattern for the afternoon: a good synthetic experiment is so clear that failure can only mean a bug — the ground truth is planted by construction. The deeper point: being unable to design one usually means you don't yet understand why your method should work, which is the real problem to solve.

SPEAKER NOTES — Slide 39 (Tests you can re-run) The operational answer to "we don't know enough": you may lack field-wide laws, but you can pin behavior down in controlled synthetic settings and keep it pinned. This is exactly what the breakout notebook will do.

SPEAKER NOTES — Slide 40 (Breakout / deliverable) This is the point of the day: students build their own synthetic-experiment notebook that justifies + tests one design choice. Tie it back to slide 10's "every choice needs a why" and the "think stupider" slide. Preview Gerych (1:30, trustworthy-by-construction) and the 4:15 share-out.

SPEAKER NOTES — Slide 41 (References) Appendix: full citations for every model in the autoregressive-EHR timeline, plus Epic's Cosmos / "Curiosity" foundation-model paper.