Research Project

EvoClinician: Self-Evolving System for White-Box Multimodal Clinical Reasoning

Jun 2026 – Sep 2026, Microsoft AI

Designed EvoClinician, a self-evolving framework that builds fully interpretable and auditable medical intelligence from multimodal EHR data.
Enabled autonomous co-evolution of a perception module that dynamically acquires missing features from the relevant modalities and a cognition module that refines its decision graph to exploit them for more accurate clinical predictions.
Achieved state-of-the-art performance over EHR foundation models and generic baselines across diverse clinical tasks while remaining fully interpretable.

2024 – 2026

Developed MicroDiffuse3D, the first volumetric restoration foundation model for 3D microscopy imaging.
Addressed the challenging 16-fold volumetric sparsity super-resolution task.
Achieved state-of-the-art performance across multiple benchmark datasets.

2025 – 2026

Developed Evo2HiC, a HiC-regularized distillation model of Evo2 that integrates long-context genomic modeling with HiC structural information into a compact and efficient CNN encoder.
Enabled a wide range of downstream applications, including high-accuracy epigenomic prediction, interpretable motif identification, and cross-species HiC resolution enhancement for 177 species.

May 2024 – Aug 2024, GE HealthCare

Developed a multi-agent system for single cell analysis, enabling low-code solutions in the field.
Streamlined programming into requirement parsing, API assignment, and parameter selection, each handled by an LLM group.
Improved the accuracy of API invocation and parameter prediction over baseline agents.

Feb 2021 – Aug 2021

Developed a cross-domain recommender system to address cold-start issues.
Utilized attention-based graph neural networks, achieved significant improvement and published a paper.