Brandon Chan MSc Defence Supervisors: Parvin Mousavi, David MasloveInternal Examiner: Tamas UngiExternal Examiner: Dr. Gordon Boyd, School of MedicineChair: Juergen DingelThesis Title:Using Deep Models to Predict Sudden Episodes of Hypotension in Critically Ill PatientsAbstract
The vast quantities of data generated and collected in the Intensive Care Unit (ICU) have given rise to large retrospective datasets that are frequently used for observational studies. The temporal nature and fine granularity of much of the data collected in the ICU enable the pursuit of predictive modeling. In particular, forecasting acute hypotensive episodes (AHE) in intensive care patients has been of recent interest to researchers in the healthcare domain. Advance warning of an impending AHE may give care providers additional information to help mitigate the negative clinical impact or prompt a search for an evolving disease process. However, the currently accepted definition of AHE is restrictive and does not account for inter-patient variability.
In this thesis, a novel definition of an AHE based on patient-specific features of blood pressure recordings is proposed. Next, deep learning methods are explored to predict the onset of patient-specific episodes of hypotension at variable event horizons and amounts of observational data.
An initial study on a single institutional cohort is conducted before extending the use of deep models to an external validation cohort. Using a single-institutional cohort of 538 patients, an initial deep model was able to successfully predict the onset of an episode of hypotension with an area under the receiver operating characteristic curve (AUROC) up to 0.87. When extending the use of deep models to an institutionally external cohort of 223 patients, a refined model was able to successfully predict the onset of an episode of hypotension in the internal dataset with an AUROC up to 0.86 and the external cohort with AUROC up to 0.87. Deep models are shown to outperform a baseline logistic regression approach inspired by the current state-of-the-art on both internal and external test data. The outcome of this work presents a generalizable definition of a patient-specific event and a complimentary deep model that can successfully predict the onset of an episode of hypotension across care centers.