Transcript Greg Cooper - Department of Biomedical Informatics

Gregory Cooper
Professor of Biomedical Informatics
Director, Center for Causal Discovery
Vice Chair, Department of Biomedical Informatics
Research involves the use of
probability theory, decision
theory, Bayesian statistics,
machine learning, and artificial
intelligence to address biomedical
informatics problems.
Causal discovery
of biomedical knowledge from big data
Problem: How to discover causal knowledge from big
biomedical datasets?
Approach: Improve the efficiency of existing causal
discovery algorithms and develop new algorithms. Make
them readily available to biomedical scientists and easy to
use.
Funding: NIH BD2K U54HG008540 (Cooper, Bahar, Berg)
Query Models
View Models
Select
Data
Select
Algorithm
Perform
Causal
Analysis
Compare Models
Causal
Models
Annotate Models
Store Models
Share Models
Discovering tumor-specific drivers and
pathways of cancer
Problem: How to discover the genomic drivers of an individual
tumor?
Approach: Use an instance-specific Bayesian causal discovery
approach and implement it to run in parallel on GPUs on the
Pittsburgh Supercomputer Center.
Funding: NIH BD2K U54HG008540 (Cooper, Bahar, Berg)
Project lead: Xinghua Lu
Machine-learning-based clinical alerting
Problem: How to detect in real time a wide variety of medical errors from
data in the EMR?
Approach: Use machine learning to construct a probabilistic model of
usual care. If current care of a patient is highly unusual according to
the model, raise an alert.
Funding: NIH / NIGMS R01GM088224 (Hauskrecht, Clermont, Cooper)
Detecting and characterizing disease outbreaks using
probabilistic modeling
Problem: How to detect and characterize new outbreaks of
infectious disease in the population?
Approach: Link a probabilistic epidemiological model of outbreak
disease in the population to probabilistic models of patient
disease in emergency departments that are capturing patient
data electronically. Apply Bayesian inference.
Funding: NIH / NLM R01LM011370 (Wagner)
Predicting cancer outcomes from a combination
of clinical and omic data
Problem: How to accurately predict the outcomes (e.g., tumor
metastasis) of a patient with cancer?
Approach: Automatically construct higher-level features (e.g., cell
signaling pathways) from raw omic and clinical data. Use those
features and machine learning to predict clinical outcomes.
Funding: PA DOH CURE grant [forthcoming] (Cooper, Bar-Joseph)