Research
Our group studies how medical AI can learn from human expertise and become useful within the conditions of real care. Four connected questions organize this work.
Learning from clinical expertise
Clinical expertise is more than a store of facts. A useful agent must know what information to seek, which evidence should change its conclusion, how to learn from experience, and when to draw on another specialist.
- MedGuideX turns executable guideline logic into verifiable training supervision, helping models learn clinical decision processes rather than retrieve more text.
- GuideSkill represents guideline knowledge as inspectable, reusable agent skills that can be updated as evidence and experience change.
- Improving Clinical Diagnosis with Counterfactual Multi-Agent Reasoning asks what would change a diagnosis, making evidence–diagnosis relationships explicit and easier for agents to challenge.
- SEA learns across a stream of cases by turning recent diagnostic experience into reusable reasoning rules.
- MedRouter learns which specialist medical models to consult and how their complementary knowledge improves a final answer.
Testing AI in realistic clinical settings
Static question answering leaves out much of clinical work. We build benchmarks and simulations that preserve hidden information, open-ended interaction, multidisciplinary discussion, imperfect evidence, and the limited time of human experts.
- MedConceal tests whether an agent can uncover concerns a simulated patient has not yet revealed and then respond effectively.
- OpenTumorBoard reconstructs real multidisciplinary cancer discussions, testing both specialist responses and full board deliberation across 611 patient cases.
- Characterizing Rhetorical Misalignment combines a clinical human-subject study with simulated decision-makers to measure when an AI response changes a correct decision into an incorrect one.
- LONGQAEVAL identifies which units of expert annotation are most reliable for different qualities of long-form clinical answers and how to reduce evaluation cost.
- Does Accuracy Equal Evidence? shows that answer accuracy can remain stable while the evidence supporting a model’s reasoning deteriorates, motivating evaluation beyond final answers.
Reasoning across time and modalities
Clinical evidence is distributed across visits, records, images, and specialist perspectives. Medical AI must preserve those relationships instead of flattening a patient’s history into disconnected facts.
- EHR-RAG preserves event type and time while retrieving evidence from long-horizon health records, rather than treating a patient’s history as flat text.
- MedRouter evaluates specialist routing across text-based and multimodal medical questions, showing why different evidence may call for different model expertise.
- OpenTumorBoard captures how longitudinal history and multimodal clinical observations are integrated through multidisciplinary discussion.
- Radiologically Relevant Clinical History Summarization studies how information from clinical notes can better support imaging protocoling and interpretation.
Supporting human decisions and communication
Medical information only helps when it fits the knowledge, goals, and decision context of the person using it. We study both how AI changes human judgment and how communication can be adapted while preserving accuracy and trust.
- Characterizing Rhetorical Misalignment shows that a factually correct model can still induce harmful clinical decisions through how it presents information.
- ReLay evaluates whether personalized plain-language summaries improve understanding and what tradeoffs personalization can introduce.
- Personalized jargon identification models which terms are unfamiliar to a particular reader rather than assuming one universal definition of jargon.
- APPLS tests whether automated metrics for plain-language summarization reflect the qualities people actually value.
See Google Scholar for citation counts and additional records.
Preprints
- MedRouter: Demystifying Knowledge Differences Across Medical LLMs for Routing-Based ReasoningarXiv2026
- OpenTumorBoard: A Real-World Benchmark of Multidisciplinary Tumor Board Discussion TrajectoriesarXiv2026
-
- GuideSkill: Evolving Executable LLM Agent Skills for Guideline-Grounded Clinical ReasoningarXiv2026
-
- MedGuideX: Internalizing Decision Logic from Executable Guidelines into Large Language Models for Clinical ReasoningarXiv2026
- EHR-RAG: Bridging Long-Horizon Structured Electronic Health Records and Large Language Models via Enhanced Retrieval-Augmented GenerationarXiv2026
Publications
2026
- Radiologically Relevant Clinical History Summarization with Large Language Models: A Multireader Performance StudyRadiology2026
- ReLay: Personalized LLM-Generated Plain-Language Summaries for Better Understanding, but at What Cost?COLM2026
- MedConceal: A Benchmark for Clinical Hidden-Concern Reasoning Under Partial ObservabilityCOLM2026
-
- LONGQAEVAL: Designing Reliable Evaluations of Long-Form Clinical QA under Resource ConstraintsEMNLP2026
- Are LLM-generated plain language summaries truly understandable? A large-scale crowdsourced evaluationJBI2026
2025
-
- Coherence and comprehensibility: Large language models predict lay understanding of health-related contentJBI2025
2024
2022
- Mixed-methods evaluation of three natural language processing modeling approaches for measuring documented goals-of-care discussions in the electronic health recordJPSM2022
2021
- Crossing the “Cookie Theft” corpus chasm: applying what BERT learns from outside data to the ADReSS challenge dementia detection taskFrontiers2021
2020
- Spatial radiation dose influence on xerostomia recovery and its comparison to acute incidence in patients with head and neck cancerAdv Radiat Oncol2020
- The performance of immunocytochemistry staining as triaging tests for high-risk HPV-positive women: a 24-month prospective studyJ Oncol2020
2019
- Dose/volume histogram patterns in salivary gland subvolumes influence xerostomia injury and recoverySci Rep2019
- Voxel dose pattern for patient-reported dysphagia among head and neck cancer patients receiving definitive radiotherapyIJROBP2019
- Low peripheral blood leukocyte counts during radiation therapy for head and neck cancer predict distant relapse and dysphagiaIJROBP2019
- Evaluation of Patient vs. Provider Based Assessments of Xerostomia and Dysphagia Compared Against Weight Loss in the Irradiated Head and Neck Cancer PatientIJROBP2019
- Definitive Management of Early Stage Oropharyngeal Carcinoma: A Comparison of Long Term Outcomes Following Transoral Surgery or Definitive ChemoradiationIJROBP2019
- A Feasibility Study of Xerostomia Outcome-Based Treatment Planning to Improve the Probability of Xerostomia Recovery in Head and Neck Cancer PatientsIJROBP2019
- An Outcome-Based Treatment Planning Strategy to Improve Xerostomia Symptom in Head and Neck Cancer PatientsMed Phys2019
2018
- Quality of life differences in male and female patients with head and neck cancerIJROBP2018
- Head and Neck Salivary Gland Spatial Dose-Volume histogram (DVH) Patterns Predict Radiation-Induced Xerostomia, From Injury to RecoveryMed Phys2018