Clinical ML has a generalization problem. The standard playbook: train a model, watch it fail at the next hospital, and retrain it with new tricks. We invert this! Don't change the modelโchange the input so any model will generalize. Introducing Record2Vec at #ICLR2026! ๐๐ 1/5
Check out our work! ๐
๐ Paper: https://t.co/QlKsl4xVkQ
๐ Project: https://t.co/vpldjQ3KMm
๐ At #ICLR2026? Come find us at poster P3-#1123!
๐๏ธ Thursday, April 23 ๐ 10:30 AM โ 1:00 PM
Come say hi! ๐๐ง๐ท #ICLR2026#HealthAI#MachineLearning#LLMs
Clinical ML has a generalization problem. The standard playbook: train a model, watch it fail at the next hospital, and retrain it with new tricks. We invert this! Don't change the modelโchange the input so any model will generalize. Introducing Record2Vec at #ICLR2026! ๐๐ 1/5
The deployment payoff: one representation, many predictors, many sites. ๐ฅโก๏ธ๐ฅ Across MIMIC-IV, HiRID, and PPICU, we see smaller performance drops under hospital transfer, stronger few-shot learning, and no added demographic leakageโall without retraining the whole stack! 4/5
๐จ Demos Deadline Extended to October 3rd, 11:59pm AoE
This is your chance to showcase mature tools that are making an impact in clinical or health settings.
Accepted demos will be presented live at #ML4H2025, with select submissions featured in talks.
https://t.co/fNPnqL10BT
1/7 ๐ Thrilled to announce that our paper ExOSITO: Explainable Off-Policy Learning with Side Information for ICU Lab Test Orders has been accepted to #CHIL2025! Please feel free to come by my poster session this Thursday to chat. #MedAI#HealthcareAI
7/7 ๐ก Huge thanks to our clinical collaborators and ICU data teams! If you're at #CHIL2025 this Thursday, drop by my poster sessionโhappy to demo ExOSITO and gather your feedback. ๐ #MachineLearning#AI4Health
Can neural networks learn to map from observational datasets directly onto causal effects?
YES! Introducing CausalPFN, a foundation model trained on simulated data that learns to do in-context heterogeneous causal effect estimation, based on prior-fitted networks (PFNs). Joint work with @Layer6AI & @hamid_R_kamkar
w/ @_valthomas, Jeremy Ma, Benson Li, Jesse C. Cresswell, & @rahulgk
๐ArXiv: https://t.co/jc9plTMo44
๐Code: https://t.co/MVO8j24mR8
๐ฃ๏ธOral paper @ ICML SIM workshop
๐งต[1/7]
๐ข New Paper at #UAI2023! ๐
Dependent censoring presents a *common and substantial* source of bias in modern survival analysis. We show how to use copulas to mitigate this bias.
https://t.co/os7Ei8xw8R
So what is dependent censoring? And why is it problematic? ๐งต๐ (1/n)
There's been a lot of success in causal effect estimation using machine learning. But what if point identification is impossible? Our NeurIPS 2022 paper, "Partial Identification of Treatment Effects with Implicit Generative Models," estimates bounds on causal effects instead. ๐งต