Tianxi Cai
Harvard University, MA·Mathematics & Statistics
Hospitals' electronic health records (EHR) hold data on millions of real patients, including far more varied people than the restricted groups enrolled in clinical trials. In principle, these records could be used to predict who will develop a disease, such as colon cancer, or respond to a treatment.
- Fields
Mathematics & Statistics
Public Health & Epidemiology
Computer Science
- Topics
electronic health records
real-world data
risk prediction
- Grants
- 1 known grant, through Apr 2027
What they work on
- Develop semi-supervised risk prediction methods that combine noisy electronic health record surrogates, such as billing codes, with a small set of chart-reviewed outcome labels, even when predictors are numerous.
- Create semi-supervised transfer learning methods that adapt risk models trained on one EHR cohort to a new population that lacks gold-standard outcome labels.
- Build distributed learning methods that fit high-dimensional prediction models across multiple EHR systems without sharing patient-level data, preserving privacy.
More about the research
In practice, the reliable outcome labels needed for such predictions require slow manual review of charts. Easy stand-ins such as billing codes are often wrong, which can bias results. Models trained at one hospital also tend to perform poorly at another, and privacy rules make it hard to pool records across sites.
Why it matters
Statistical tools that turn noisy health records into accurate, interpretable risk models could help clinicians predict disease progression and treatment response, for example in cancer. Such models could support more efficient and cost-effective personalized care and help check whether clinical trial findings hold up in broader, real-world patient populations.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Mei Liu
University of Florida, FL·Medicine & Clinical Research
Develop methods to build the best individualized cohort of similar past patients from EHR data, as the training set for a personalized AKI risk model for each new hospital patient.
Akihiko Nishimura
Johns Hopkins University, MD·Mathematics & Statistics
Develop statistical methods that combine Bayesian hierarchical models, biomedical ontologies and high-dimensional transfer learning to pool genetic and clinical data from sites that differ in patients, data coding and practice.
Roy Adams
Johns Hopkins University, MD·Computer Science·Early-career
Develop machine learning methods that combine several imperfect record-based clues for dementia while modeling the error and bias in each, to label whether a patient has Alzheimer's disease or a related dementia.
The school
Harvard University: 42 labs on PI Pages, $1.5B a year on research, 712 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.