Events

Machine Learning Methods for Clinical Risk Prediction

Biomedical Engineering
Lecture / Panel
 
Open to the Public
""
Zoom

Speaker:

Adler Perotte, MD, MA
Assistant Professor
Dept. of Bioinformatics
Columbia University Irving Medical Center

Abstract:

Predicting the risk for a patient to develop a certain disease or achieving a particular treatment outcome has become a mainstay of modern medicine. Examples include, how likely will a woman with certain genetic marker develop breast cancer; what are the chances that an obese patient will have a heart attack; what is the likelihood of a COVID-19 patient to survive if he/she is on a ventilator, etc. In this talk Dr. Adler will look at the mathematical models and methods that are currently used to come up with these predictions. What assumptions are used in formulating these models and what determines their accuracy? In particular, he will focus on three different cases: The first study will show how the combi-nation of disparate sources of data in a single model improves our ability to predict chronic kidney disease progression. The second study demonstrates how modern deep-learning methods, when applied to survival analysis, predict cardiovascular risk better than well-established risk scores. Lastly, he will examine how data available at the standard point of care, can be predicted outcomes of interest in COVID-19 patients, including the need for respiratory support, the need for kidney support, and early return to the hospital.


Dr. Perotte earned his Artium Baccalaureus (AB) from Princeton University and his MD from Columbia University Medical School. After being a research specialist working between Princeton’s computational memory and artificial intelligence laboratories, he accepted a position as a National Library of Medicine (NLM) Postdoctoral Fellow in Biomedical Informatics at Columbia University. He is now an Assistant Professor in the same department and focuses on Bayesian inference and predictions based on electronic health record data and metabolomics. Dr. Perotte has also been involved in the design and implement-tation of the infrastructure and methods for the Observation Health Data Sciences and Informatics (OHDSI) collaborative, which currently hosts more than 500 million patient records.

chart depicting depression as extracted from the OHDSI database
Predicting treatment pathways for patients with depression extracted from the OHDSI database. The inner circle shows the first relevant medication that the patient took, the second circle shows the second medication, and so forth. Only four levels are shown, but up to 20 medications were recorded.