AI for Healthcare Sihang Zeng

Editorial Atlas

Designing AI that reads the longitudinal patient record as a living clinical narrative.

From trajectory modeling to biomedical foundation models, his research translates messy real-world health data into trustworthy signals for screening, prognosis, and clinical decision-making.

View Publications See Global Reach
2026 Expected PhD completion
14 Scholar-listed works
3 Core research threads
Portrait of Sihang Zeng
Current base Seattle, Washington

PhD Candidate in Biomedical Informatics, University of Washington

About

Clinical AI with a longitudinal view

Sihang Zeng is a PhD candidate in Biomedical Informatics at the University of Washington, advised by Ruth Etzioni and Meliha Yetisgen. His work spans longitudinal EHR modeling, clinical large language models, biomedical representation learning, and evaluation frameworks for healthcare AI. Before UW, he studied Electronic Engineering with a minor in Statistics at Tsinghua University. His public academic profile also highlights earlier work on biomedical term representation and knowledge graph construction, forming a throughline from biomedical language understanding to patient-centered AI systems.

Current Lens

AI for healthcare that respects time, uncertainty, and care context

The throughline is simple: patient records are not isolated events. They are sequences, dependencies, missed opportunities, and evolving signals. The research aims to model them that way.

Research

Three threads run through the work

The site is organized around longitudinal patient narratives, clinical foundation models, and evaluation systems that can earn trust in real biomedical settings.

01

Longitudinal Patient Modeling

Learning patient-state trajectories from sparse, irregular EHR histories to support risk prediction, prognosis, and clinical foresight.

  • EHR
  • Temporal ML
  • Representation Learning
02

Foundation Models for Clinical Data

Extending language-model style reasoning into multimodal, longitudinal healthcare data while preserving clinical grounding and decision relevance.

  • LLMs
  • Generative Models
  • Healthcare AI
03

Trustworthy Evaluation

Building benchmarks and evaluation strategies that make biomedical and scientific AI more reliable, transparent, and useful in practice.

  • Evaluation
  • Clinical Safety
  • Scientific Review

Selected Work

Recent papers shaping the narrative

A few representative projects spanning trajectory reasoning, biomedical generalists, and evaluation systems.

NeurIPS 2025 GenAI4Health 2025

Traj-CoA: Patient Trajectory Modeling via Chain-of-Agents for Lung Cancer Risk Prediction

Sihang Zeng, Yujuan Fu, Sitong Zhou, Zixuan Yu, Lucas Jing Liu, Jun Wen, Matthew Thompson, Ruth Etzioni, Meliha Yetisgen

Frames longitudinal risk prediction as a coordinated agent workflow over patient trajectory evidence.

  • Longitudinal EHR
  • Agentic Systems
  • Cancer Screening
Paper
Machine Learning for Healthcare 2025 2025

TrajSurv: Learning Continuous Latent Trajectories from EHR for Trustworthy Survival Prediction

Sihang Zeng, Lucas Jing Liu, Jun Wen, Meliha Yetisgen, Ruth Etzioni, Gang Luo

Learns smooth latent patient trajectories that improve survival prediction while preserving interpretability.

  • Survival Analysis
  • Trustworthy ML
  • Patient Representation
Paper Code
EMNLP 2025 2025

ReviewRL: Towards Automated Scientific Review with RL

Sihang Zeng, Kai Tian, Kaiyan Zhang, Yuru Wang, Junqi Gao, Runze Liu, Sa Yang, Jingxuan Li, Xinwei Long, Jiaheng Ma, Biqing Qi, Bowen Zhou

Explores reinforcement learning for more grounded, analytically useful automated scientific reviews.

  • Scientific AI
  • RL
  • Evaluation
Paper
NeurIPS 2024 Datasets and Benchmarks 2024

UltraMedical: Building Specialized Generalists in Biomedicine

Kaiyan Zhang, Sihang Zeng, Ermo Hua, Ning Ding, Zhang-Ren Chen, Zhiyuan Ma, Haoxin Li, Ganqu Cui, Biqing Qi, Xuekai Zhu, Xingtai Lv, Jin-Fang Hu, Zhiyuan Liu, Bowen Zhou

Builds broadly useful biomedical models without sacrificing domain specialization.

Spotlight
  • Biomedical LLMs
  • Generalist Models
  • Benchmarking
Paper Code

Timeline

A recent arc of papers, milestones, and field wins

Recent work moves across clinical modeling, scientific evaluation, and biomedical foundation-model building, while staying grounded in real healthcare impact.

  1. Jan 2026
    paper

    MARTI accepted at ICLR 2026

    A new step toward agentic reasoning for longitudinal clinical modeling.

    Open source
  2. Dec 2025
    milestone

    Passed PhD general exam

    Advanced to PhD candidacy in biomedical informatics.

  3. Oct 2025
    award

    1st Place at ChemoTimelines 2025 Challenge

    Oral presentation at Clinical NLP for timeline-centric oncology reasoning.

  4. Oct 2025
    paper

    Traj-CoA accepted at NeurIPS 2025 GenAI4Health

    Chain-of-agents modeling for lung cancer risk prediction from patient trajectories.

    Open source
  5. Oct 2025
    career

    Joined Truveta as ML Intern

    Applied foundation-model thinking to large-scale clinical data in industry.

  6. Aug 2025
    paper

    ReviewRL accepted at EMNLP 2025

    Reinforcement learning for scientific reviewing workflows.

    Open source
  7. Jul 2025
    paper

    TrajSurv accepted at MLHC 2025

    Continuous latent trajectories for trustworthy survival prediction.

    Open source
  8. Sep 2024
    paper

    UltraMedical spotlight at NeurIPS 2024 Datasets and Benchmarks

    Specialized generalist models for biomedicine.

    Open source
  9. Sep 2023
    career

    Started PhD at the University of Washington

    Began doctoral work in biomedical informatics and healthcare AI.

Affiliations

Built across academic, translational, and industry settings

The work sits at the intersection of biomedical informatics, cancer prevention, and large-scale real-world health data.

University of Washington
Department of Biomedical Informatics and Medical Education
Fred Hutchinson Cancer Center
Truveta
Tsinghua University

Global Reach

A live, privacy-conscious map of the site's global readership

Each point represents anonymized aggregate visitor activity from the site. If the API is not configured yet, the globe still renders and explains how to activate live tracking.

Total visits
--
Countries
--

Connecting to the visitor atlas.