ICLR 2027 Workshop on Human-Centered AI for Health Multimodal Health Intelligence, Self-Evolving Agents,
and Human-Centered Evaluation

April 29, 2027San Francisco, California, USA

Keynote Speakers TentativeHuman-Centered AI for Health @ ICLR 2027

Talk titles and speaking order will be announced.

Mihaela van der Schaar

University of Cambridge

Suchi Saria

Johns Hopkins University & Bayesian Health

Agenda TentativeHuman-Centered AI for Health @ ICLR 2027

One-day program on April 29, 2027. Each keynote slot lasts 15 minutes including Q&A; speaker order and talk titles will be finalized with the speakers.

Morning Session 08:00-12:30

08:00-08:10

Welcome and opening remarks

08:10-09:10

Keynotes I: Multimodal and multisensory health intelligence

09:10-09:40

Contributed Talks I

09:40-10:00

Coffee break and informal discussion

10:00-11:00

Keynotes II: Self-evolving agentic systems and learning from experience

11:00-11:45

Panel I: Human factors in health AI

11:45-12:30

Interactive posters and demos I

12:30-13:30

Lunch and informal exchange

Afternoon Session 13:30-17:00

13:30-14:30

Keynotes III: Reliability, safety, and human-centered evaluation

14:30-15:00

Contributed Talks II

15:00-15:20

Coffee break and informal discussion

15:20-16:05

Panel II: Evidence of patient benefit

16:05-16:50

Interactive posters and demos II

16:50-17:00

Paper awards, discussion summary, and closing

Panel Discussions

Each 45-minute panel uses a concrete care scenario to guide 30 minutes of moderated discussion and 15 minutes of audience questions. Panelists will be announced.

Panel I

Human factors in health AI

How should health AI respond when patient preferences, clinician judgment, or changing care circumstances challenge its assumptions?

Panel II

Evidence of patient benefit

When does better AI performance translate into better care for patients, and what is missing when it does not?

With speaker consent, invited talks and panels will be recorded and livestreamed, and posted on this website after the event.

Call for PapersHuman-Centered AI for Health @ ICLR 2027

Submissions DueFebruary 1, 2027

NotificationFebruary 26, 2027

Final ProgramApril 14, 2027

WorkshopApril 29, 2027

Submit on OpenReview opens soon All deadlines are 11:59 PM AoE (Anywhere on Earth).

About

AI for health matters only if it works for the people it serves: patients, clinicians, caregivers, people managing their own health, and public-health practitioners. Capability alone does not establish benefit. Stronger models have not, on their own, improved clinical decisions, and fairness measured on a model does not guarantee fairness for the people affected. This workshop involves those people in defining what AI should do, how it should interact, and how it should be evaluated.

Multimodal foundation models and agentic systems expand what AI can do for health, from interpreting heterogeneous data to supporting decisions and learning through interaction. Their value depends on whose needs define the task, how responsibility is shared among developers, practitioners, and users, and how systems incorporate feedback while retaining human oversight as they adapt. We bring together machine learning, health, human-computer interaction, and AI safety around three questions: what evidence represents a person's needs, how AI can act and adapt under human guidance, and when reliance on a system is justified.

The workshop aims to

Connect multimodal learning to people's needs
which modalities and measurements improve health decisions, and how collection burden, privacy, and accessibility shape them.
Make adaptation useful and controllable
how agents learn from feedback and delayed outcomes while people can inspect, correct, or stop them, and harmful updates can be reverted.
Build complementary evaluation strategies
what benchmarks, simulations, expert assessment, participatory studies, and prospective evidence each reveal or miss.
Produce a shared research agenda
a public synthesis of open questions and evaluation priorities covering decision quality, workload, appropriate reliance, and health benefit.

Topics of Interest

We welcome methodological, empirical, and conceptual contributions in three thematic areas. Submissions may address a single area or connect several.

Work on individual modalities, conventional predictive models, datasets, benchmarks, negative results, failure analyses, and human factors is welcome when it advances the questions below. Applications span clinical care, mental health and computational psychiatry, aging, community and public health, and everyday wellbeing.

Multimodal and Multisensory Health Intelligence

Learning from health records, clinical text, imaging, physiological signals and time series, wearables, speech, video, egocentric, and environmental observations; alignment across incomplete, irregular, or conflicting modalities; and longitudinal, context-aware, personalized patient representations. The human-centered question is what evidence represents a person's needs.

Multimodal fusion and cross-modal alignment across EHR, text, imaging, time series, wearables, speech, and video
Handling incomplete, irregular, or conflicting evidence across time and modalities
Longitudinal, context-aware, and personalized patient representations
Foundation models for clinical and consumer health data, including sensor-language alignment
Egocentric, ambient, and environmental sensing for health and daily living
Which measurements are useful to a person, and when acquiring more information is worth the burden
How collection burden, privacy, and accessibility shape model design
How care and measurement processes shape the data available about each patient
Representations that account for diverse populations and contexts

Self-Evolving Agentic Systems and Learning from Experience

Healthcare agents that retrieve evidence, reason over incomplete histories, plan, use tools, and coordinate with other agents, and that improve through continual, active, and reinforcement learning by updating parameters, memory, prompts, tools, or policies. The human-centered question is how AI can act and adapt under human guidance.

Retrieval, reasoning, planning, and tool use over incomplete patient histories
Multi-agent coordination in clinical, patient-facing, and administrative workflows
Deciding what information to acquire, when to act, and when to defer to or seek human input
Continual, active, and reinforcement learning from clinician feedback, interaction, and delayed outcomes
Self-evolving agents that update memory, prompts, tools, policies, or parameters
Learning from experience, synthetic data, and simulation, including simulated clinical environments
Preserving prior capabilities and respecting safety constraints while limiting data and compute
Human-in-the-loop decision support, shared decision-making, and patient communication
Care navigation, behavior-change support, and home-based or conversational assistance
Letting people inspect, correct, override, or stop delegated actions

Reliability, Safety, and Human-Centered Evaluation

Causal and probabilistic inference, uncertainty quantification, robust learning under distribution shift, privacy, fairness, and post-deployment monitoring, together with evaluations that combine automated benchmarks, realistic synthetic scenarios, expert assessment, participatory user studies, and prospective evidence. The human-centered question is when reliance on a system is justified.

Causal and probabilistic methods for assessing clinical benefit, harm, and reliability
Uncertainty quantification, calibration, and uncertainty communication to users
Robustness under distribution shift and reliability across patient groups, care settings, and time
Fairness, privacy, and equitable access connected to the people and settings affected
Independent prospective studies, stress tests, and red teaming
Benchmarks, datasets, negative results, and failure analyses
Post-deployment monitoring, safe updates, and rollback of models, memory, or tools
Explanation, evidence inspection, and error correction interfaces
Appropriate reliance, automation bias, and human oversight
Participatory studies with clinicians, patients, and caregivers
Outcomes beyond accuracy: decision quality, workload, accessibility, quality of life, and health benefit
Accessible and multilingual interfaces for wearable, ambient, conversational, embodied, or home-based systems

Submission Formats

Three non-archival formats across all thematic areas. Limits exclude acknowledgments, references, and appendices.

Research Papers

Up to 8 pages

Methods or empirical findings with supporting evidence, including short research papers. A clearly stated problem, a described method, appropriate baselines, and evaluation that supports the claims. Work in progress is welcome when current results support the central claim.

Demonstration and Position Papers

Up to 5 pages

Either a functioning system, deployed, in pilot, or in working prototype state, with its intended users and evidence that it works as described (screenshots, interactions, a video, usage data, or user feedback); or a clear position on health AI research or practice, supported by evidence and engaging seriously with counterarguments.

Tiny Papers

Up to 2 pages

Focused early-stage contributions that make one point well: a new idea with a clear hypothesis, a replication in a health setting, or a re-analysis of existing data or benchmarks. A lighter entry point for new and under-resourced researchers; must be primarily human-written.

Review criteria

Soundness

rigor of the method, system, or argument and its evidence.

Novelty

of the method, system, benchmark, dataset, idea, or findings.

Relevance and impact

for health stakeholders and the workshop themes.

Clarity

of presentation, including a discussion of limitations.

All accepted papers are presented in the interactive poster and demo sessions; contributed talks are selected from them. Three Outstanding Paper Awards, one per format, are announced at the closing session.

How to Submit

Format

ICLR 2027 LaTeX style with \usepackage{iclr2027_conference} and without \iclrfinalcopy. Supplementary material (code, data, videos) goes in appendices.

Anonymity

double-blind. No author names, affiliations, identifying acknowledgments, or non-anonymized links; violations may lead to desk rejection.

Eligibility

original, unpublished work and work in progress. Preprints such as arXiv are permitted; work already published at an archival venue is not. Off-topic, duplicate, and abstract-only submissions are screened out.

Platform

OpenReview (link to be posted here). Register early; new account approval can take several days.

Reviews

at least two human reviews per paper, matched by topic and format, with at most four papers per reviewer. No rebuttal stage, so make the paper self-contained.

Conflicts

submissions involving an organizer or their institution are handled by a non-conflicted committee member.

AI assistance

permitted for authors and reviewers, who remain responsible for verifying its output and protecting confidential material. Submissions and reviews must be primarily human-written; AI-generated tiny and short papers are not permitted.

Publication Policy & FAQ

This workshop is non-archival: no proceedings, no formal publication, and you keep full copyright. Accepted papers are made public on OpenReview and imported into iclr.cc by February 26, 2027, with links posted here.

Can I submit work that is under review elsewhere, was withdrawn from a main track, or send my paper elsewhere afterwards?

Yes to all three, because the workshop is non-archival. Concurrent submission requires that the other venue's dual-submission and anonymity rules permit it, and some venues restrict prior presentation even at non-archival workshops; checking this is the authors' responsibility.

Can I submit work that has already been published or accepted elsewhere?

Work already published in an archival venue is not eligible. Work accepted but not yet published may be submitted only if the other venue permits it. Non-archival preprints are fine.

Will my accepted paper be public?

Yes. Accepted papers are posted publicly on OpenReview and imported into iclr.cc by February 26, 2027, in line with ICLR 2027 workshop requirements.

What is a Tiny Paper, and do page limits include references?

A two-page format for focused early-stage contributions such as a new idea, a replication, or a re-analysis. For all formats, acknowledgments, references, and appendices are excluded from the limit, and reviewers are not required to read appendices. No ICLR checklist or reproducibility statement is required.

Do I have to present in person?

The workshop is held in person in San Francisco, and at least one author of each accepted paper is expected to present there. With speaker consent, talks and panels will be recorded and posted on this website for those who cannot attend.

Can I use AI tools when writing my paper?

AI assistance is permitted, but authors remain responsible for verifying its output and for the content. Submissions must be primarily human-written; AI will not serve as a primary author, and AI-generated tiny and short papers are not permitted.

Can the author list change after submission?

No authors may be added after the deadline at any stage. Reordering is permitted for accepted papers with every author's consent. Removing an author requires written consent from every author, including the one removed, sent to the organizers with a justification.

Other questions? Email the organizers at xhan56@jh.edu or gauthamkrishna@utexas.edu.

OrganizersHuman-Centered AI for Health @ ICLR 2027