Machine Learning for Health Care.
Posted on :: Tags: conference, ml, machine learning, medical, health care, AI

Table of Contents

TL;DR: Biased memos from MLHC 2026 talks/posters.

Thu Aug 13 09:00 EDT 2026

  • Google AMIE team

    • Diagnosis: https://www.nature.com/articles/s41586-026-10764-5

    • Expert-level medical care https://www.nature.com/articles/s41586-026-10764-5

    • Had access to NEJM case studies and single-site EHR text

    • No access to researchers

  • How can we leverage RCT in observational data analysis or vice versa

  • The Illusion of Learning from Observational Data: An Empirical Bayes Perspective

  • Many people seem to (or would like to) train LLMs on PT-HCP interaction data

  • Having an agentic system for healthcare, are we opening up the possibility of over-triage, over-diagnosis, and over-treatment?

  • No physician will be judged numerically.

  • What will be the major obstacles for human-computer collaborations?

Thu Aug 13 14:05 EDT 2026

  • Ilya Shpitser

    • "Recursive equations for imputation of missing not at random"
    • Two approaches:
      • Model missing status, e.g., (1) inverse propensity weighting (2) semi-parametric
      • Model missing values
    • Missing data = a counterfactual problem
    • MCAR: $X^{(1)} \to X \gets R$,
    • MAR: confounder $\to X^{(1)}$ and $R$
    • MNAR: $X^{(1)} \to R \to X$ and $X^{(1)} \to X$, "self-censoring"
    • Need to model $p(X|R, X^{(1)})$
      • No self-censoring model, a conditional MRF
      • IF a DAG is a submodel of NSC...
    • Semi-parametric estimator available
    • MICE relies on MAR
    • goal: $p(X^{(1)},O|r)$ = extrapolation $\times$ interpolation
    • extra: $p(X|X',O,r)$, Gibbs
    • intra: $p(X'|O,r)$
    • Estimable vs. estimation

  • Concept bottleneck model

    • Can we prevent a large network from forgetting in continual learning?
    • Just because we can choose not to learn a part of a model... we switch off gradient flows... (soft or hard)?

  • Reinforcement learning with data censorship

    • Matt Engelhard, Duke
    • POMDP; Survival data sets are clearly censored.
    • Early alert, pseudo label, post-hoc adaptation

Fri Aug 14 09:50 EDT 2026

  • Anna Goldenberg

    • Heterogeneity in large-scale ICU data
    • Self-supervised learning representation learning
    • Different sampling intervals
    • Not high-dimensional?
    • Hierarchical Dirichlet Process, flow model...
      • Representation = latent states?
    • How do we take into account patient specificity?
    • DynaSub: adaptive subgrouping with encoders

  • Hoifung Poon

    • Learning the language of patients across modalities
    • Virtual patient
    • Targeted therapy worked but hard to deal with relapse
    • Immunotherapy, Keytruda (checkpoint blockade), not for everyone
      • Need to learn tumour microenvironment grammar
      • The other side... autoimmune disease
    • Amara's Law

      "We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run"

    • Physical experiments are expensive
    • Can we simulate a clinical trial based on real-world data? (NEJM AI)
    • Patient journey is multimodal
    • Precision health is a multimodal generation problem
      • Dealing with missing information is a genuinely important challenge
      • Can we model out missing information?
      • Virtual patient, digital twin
    • Unimodal: encoder $\to$ decoder
    • Digital pathology: a transformer may not be the right architecture
      • 16 x 16 = 56 million tokens
    • A whole-slide foundation model
    • Proposing "text" as an interlingua medium
    • Spatial proteomics/transcriptomics is expensive
    • GigaTIME
    • Can we predict next medical event of a person?
      • EPIC COSMOS database contains 115 Billion medical events