I’m an Assistant Professor of Biomedical Data Science and of Medicine at Stanford University. My research focuses on training and evaluating foundation models for healthcare and is positioned at the intersection of computer science, medical informatics, and hospital systems. Much of my work explores using electronic health record (EHR) data to contextualize human health, leveraging longitudinal patient information to inform model development and evaluation. My work has appeared in NeurIPS, ICLR, AAAI, Nature Communications, and npj Digital Medicine.
Download my Research Statement
I’m recruiting postdoctoral fellows with PhDs in machine learning–related fields (e.g., computer science, data science, biomedical informatics) who are excited to work with real clinical data in close collaboration with clinicians.
If you’re interested, please email me your (1) research statement, (2) CV, and (3) the names of three references who can provide letters of recommendation.
Ideal candidates will:
- Have strong ML foundations with publications in top ML/medical AI venues (e.g., NeurIPS, ICLR, Nature Medicine, npj Digital Medicine, NEJM AI)
- Have experience with healthcare or clinical ML using real-world data
- Be excited to mentor students, collaborate with clinicians, and contribute to grants
Research Interests
Evaluating Foundation Models Reproducibility in healthcare AI is hampered by a lack of standardized benchmarks and challenges in sharing patient data, undermining the research community’s shared understanding of state-of-the-art methods. Private foundation models exacerbate this challenge by introducing additional non-reproducible components. To tackle these challenges and promote reproducibility, we’ve released new EHR datasets INSPECT, EHRSHOT, MedAlign, FactEHR, and made our foundation models available via Hugging Face
Training Multimodal Foundation Models Future healthcare models must integrate diverse data modalities, including imaging, omics, wearables, and medical literature, to capture health progression over time. Longitudinal EHRs provide critical temporal context but are noisy, requiring data-centric AI methods like cleaning, valuation, and curation. My research explores methods to transform EHR timelines into supervision sources to train robust, scalable multimodal models that better capture long-term disease progression.
Human-AI Teaming Today’s preference alignment methods only capture a coarse sense of tacit knowledge —expertise that is contextual, embedded in practice, and rarely documented. Tacit knowledge is a defining feature of complex, multi-stakeholder decision-making processes in medicine, such as tumor boards and care coordination. Successful human-AI teaming in healthcare will require new methods to study, capture, and integrate tacit knowledge into the next generation of healthcare foundation models.