Multimodal Representation Learning for Pregnancy Care

Machine learning for healthcare promises to improve our diagnostic ability, to plot optimal treatment plans, and to smooth clinical operations and lower costs. But, central to these promises is a strong, generalizable numerical representation of a patient’s prior medical history that is reflective of clinical similarity and amenable to modeling. Learning such a representation is difficult as health data, including EHR data, is notoriously messy, sparse, and multimodal (e.g., comprised of many different types of information streams). In this work, I will build a machine learning system capable of learning a suitable representation of a patient, principally by leveraging massively-multitask modeling and structured methods such as graph networks. This representation will be evaluated through its applicability in pregnancy care, specifically through its ability to perform generalizable risk profiling across key outcome measures, such as low birth-weight and pre-term birth, and through its ability to perform patient subtyping.

Faculty Supervisor:

Marzyeh Ghassemi

Student:

Partner:

Massachusetts Institute of Technology

Discipline:

Computer science

Sector:

Education

University:

University of Toronto

Program: