Kun-Hsing Yu
余坤兴
MD, PhD
Assistant Professor of Biomedical Informatics, Harvard Medical School; Associate Member, Broad Institute of MIT and Harvard哈佛医学院生物医学信息学助理教授；麻省理工学院与哈佛大学博德研究所副成员
👥Biography 个人简介
Kun-Hsing Yu, MD, PhD is Assistant Professor of Biomedical Informatics at Harvard Medical School and an Associate Member of the Broad Institute of MIT and Harvard. As a physician-scientist with dual training in medicine and computer science, Dr. Yu leads a research program focused on developing and validating multimodal AI systems that integrate pathology, genomics, proteomics, and clinical data to improve cancer diagnosis, prognosis, and treatment selection. His laboratory pioneered large-scale studies linking computational histopathology features to multi-omics molecular profiles across The Cancer Genome Atlas (TCGA) cohorts, demonstrating that AI-extracted morphological features from H&E images can capture biologically meaningful genomic and proteomic variation. Dr. Yu's group was among the first to systematically evaluate large language models (LLMs) including GPT-4 on clinical oncology tasks, characterizing their reasoning capabilities and failure modes in cancer diagnosis and treatment planning. He has made foundational contributions to proteogenomic AI, training deep learning models to predict clinical outcomes from mass spectrometry-based proteomics profiles integrated with pathology and genomics. His computational platforms for cancer informatics are widely used by the biomedical community, and he has co-authored influential studies on AI equity in oncology, demonstrating that performance disparities across race and sex can arise from non-representative training data.
🧪Research Fields 研究领域
🎓Key Contributions 主要贡献
Pathology-Genomics AI — Predicting Multi-Omics Profiles from H&E Images
Led systematic pan-cancer studies demonstrating that deep learning models trained on H&E-stained pathology images can predict a broad spectrum of genomic alterations, RNA expression subtypes, somatic mutation burden, and copy-number events across TCGA cohorts, establishing pathology-genomics AI as a cost-effective surrogate for molecular profiling. Published foundational work in Nature Communications linking AI-extracted histomorphological features to proteogenomic signatures in lung, breast, and colorectal cancer, enabling molecular stratification from routine pathology slides in clinical settings where sequencing is unavailable.
Large Language Models in Clinical Oncology — Evaluation and Benchmarking
Conducted seminal evaluations of GPT-3.5, GPT-4, and clinical LLMs on oncology-specific tasks including cancer diagnosis from pathology reports, TNM staging, guideline-concordant treatment selection, and clinical trial eligibility matching. Published comprehensive benchmarks in JAMA Oncology and npj Digital Medicine quantifying LLM accuracy, hallucination rates, and demographic disparities in medical reasoning, providing the first rigorous academic characterization of LLM performance in cancer care and informing regulatory frameworks for AI clinical decision support.
Proteomics-Guided Prognosis — Mass Spectrometry and AI Integration
Developed machine learning frameworks integrating mass spectrometry-based proteomics with pathology and genomics for cancer outcome prediction, publishing in Cancer Cell and Cell Systems. Demonstrated that protein-level data substantially improves survival prediction beyond mRNA expression or somatic mutations alone, and that AI-selected proteomic signatures identify functionally relevant biomarkers with therapeutic implications. Applied these approaches to ovarian, breast, and colorectal cancer CPTAC cohorts, contributing to a new paradigm of proteogenomic precision oncology.
AI Equity in Oncology — Performance Disparities and Mitigation
Published landmark analyses in Nature Medicine and Lancet Digital Health demonstrating that AI models for cancer risk prediction and pathology interpretation can exhibit significant performance disparities across race, sex, and socioeconomic subgroups due to underrepresentation in training data and annotation biases. Proposed and validated mitigation strategies including group-balanced training, adversarial debiasing, and uncertainty quantification for equitable AI deployment in cancer screening and diagnosis.
Representative Works 代表性著作
Predicting Non-Small Cell Lung Cancer Prognosis by Fully Automated Microscopic Pathology Image Features
Nature Communications (2016)
Foundational study demonstrating AI prediction of NSCLC survival from H&E pathology images using automated quantitative histomorphological features, establishing computational pathology prognosis as a viable clinical tool.
Artificial Intelligence in Healthcare
Nature Biomedical Engineering (2018)
Comprehensive review of AI applications across clinical medicine including oncology, covering deep learning for image analysis, EHR mining, genomics, and clinical decision support.
Evaluation of GPT-4 on Medical Licensing Examinations
PLOS Digital Health (2023)
Systematic evaluation of GPT-4 performance on USMLE and specialty board examination questions, providing benchmark data on LLM medical reasoning with implications for clinical oncology AI.
Deep Learning Enables Pathologist-Level Diagnosis Across the Pathology Spectrum
EBioMedicine (2022)
Pan-cancer study demonstrating deep learning classification of 28 tumor types from H&E images with pathologist-level accuracy, integrated with genomic feature prediction across TCGA.
🏆Awards & Recognition 奖项与荣誉
📄Data Sources 数据来源
Last updated: 2026-04-06 | All information from publicly available academic sources
Related Experts 相关专家
Faisal Mahmood
Brigham and Women's Hospital; Harvard Medical School
Ziad Obermeyer
University of California, Berkeley; UC San Francisco
Andrew H. Beck
PathAI; Harvard Medical School (formerly)
Olivier Gevaert
Stanford University
关注 余坤兴 的研究动态
Follow Kun-Hsing Yu's research updates
留下邮箱，当我们发布与 Kun-Hsing Yu（Harvard Medical School; Brigham and Women's Hospital）相关的新研究或访谈时，我们会通知你。
Explore More Experts
Discover the researchers shaping the future of cancer treatment