Multi-institute domain adaptation by adversarial constrained medical time series representation learning

Hospitals strive to perform cutting edge medical treatment, treat all patients fairly, and reduce operating costs, while also enabling caregivers to spend more time interacting with patients. Artificial intelligence and machine learning promise these things. However, medical data provides unique challenges for machine learning. Currently, if a hospital wants to include an algorithm for automated decision making, they must either secure approval to collect additional patient data or change their care practices to replicate those at other institutions. This work proposes a novel application of artificial intelligence in medicine that creates a numeric representation of patients’ electronic medical records which is constrained to be similar across all hospitals despite each hospital having different underlying operating procedures. As a result, we can directly transfer algorithms which have proven to improve care at one hospital to another, without the need for additional data collection. This research has the potential to save lives of patients who otherwise might have been overlooked, improve patient quality of life, and set a precedent for quality healthcare globally within the next three years.

Faculty Supervisor:

Marzyeh Ghassemi;Anna Goldenberg

Student:

Bret Nestor

Partner:

Vector Institute

Discipline:

Computer science

Sector:

Professional, scientific and technical services

University:

University of Toronto