A method to detect and represent temporal patterns from time series data and its application for analysis of physiological data streams
MetadataShow full item record
In critical care, complex systems and sensors continuously monitor patients’ physiological features such as heart rate, respiratory rate thus generating significant amounts of data every second. This results to more than 2 million records generated per patient in an hour. It’s an immense challenge for anyone trying to utilize this data when making critical decisions about patient care. Temporal abstraction and data mining are two research fields that have tried to synthesize time oriented data to detect hidden relationships that may exist in the data. Various researchers have looked at techniques for generating abstractions from clinical data. However, the variety and speed of data streams generated often overwhelms current systems which are not designed to handle such data. Other attempts have been to understand the complexity in time series data utilizing mining techniques, however, existing models are not designed to detect temporal relationships that might exist in time series data (Inibhunu & McGregor, 2016). To address this challenge, this thesis has proposed a method that extends the existing knowledge discovery frameworks to include components for detecting and representing temporal relationships in time series data. The developed method is instantiated within the knowledge discovery component of Artemis, a cloud based platform for processing physiological data streams. This is a unique approach that utilizes pattern recognition principles to facilitate functions for; (a) temporal representation of time series data with abstractions, (b) temporal pattern generation and quantification (c) frequent patterns identification and (d) building a classification system. This method is applied to a neonatal intensive care case study with a motivating problem that discovery of specific patterns from patient data could be crucial for making improved decisions within patient care. Another application is in chronic care to detect temporal relationships in ambulatory patient data before occurrence of an adverse event. The research premise is that discovery of hidden relationships and patterns in data would be valuable in building a classification system that automatically characterize physiological data streams. Such characterization could aid in detection of new normal and abnormal behaviors in patients who may have life threatening conditions.