Physiological time series retrieval and prediction with locality-sensitive hashing
Year: 2017
Type: dissertation
Abstract: The amount of time series data collected in the medical community has recently been exploding due to widespread affordable sensors and storage devices.However, while the massive repositories of such physiological time series data provide enormous opportunities for machine learning to make significant impacts, they are largely under-utilized due to their granular detail, overwhelming size, and lack of proper tools.Besides scale, fast yet accurate processing of physiological waveform data is desired in medical practice, especially in time-critical settings such as the intensive care unit (ICU).Efficiently leveraging these massive datasets is a key challenge that, when resolved, will support a new paradigm of scientific discovery and operational innovation in medicine.In this thesis, we develop highly efficient similarity-based methods that make it practical to search massive physiological time series repositories to rapidly identify waveforms similar to those from a given individual.We call this concept "patients with trajectories like mine."Our goal is to exploit rapid similar waveform retrieval to enable critical event prediction in the ICU setting.In order to achieve this goal, we propose to apply locality-sensitive hashing (LSH), which supports a very fast approximate nearest neighbor search in high dimensions.We empirically demonstrate that LSH based retrieval and prediction methods vastly speed up querying time while sacrificing only a trivial amount of accuracy as a cost.Despite being fast and accurate, the generic LSH has two shortcomings.First, it is capable of utilizing only one similarity measure at a time.To overcome this limit, we introduce Stratified LSH (SLSH) which finds similarity among the data from a more integrated perspective by employing multiple distance metrics in one framework.SLSH is essentially a dual-level hierarchical LSH where each LSH layer is associated with a distinct distance metric capturing a unique facet of similarity.The second shortcoming and the main bottleneck of the generic LSH is that it involves exhaustive distance calculations as a subroutine when short-listing the candidate set to find the final nearest neighbors.To surmount this, we propose Collision Frequency LSH (CFLSH) which short-lists the candidate set by simply counting the frequency I am greatly indebted to many people around me at MIT.I can never thank them enough.First of all, I would like to express my sincere gratitude to my research advisor Dr. Una-May O'Reilly whose guidance and support were unwavering these past five years.She was always positive in the face of my numerous failures along the way and encouraged me to pursue a broad vision and freely explore my own research interests.Without her passion,
Author: Yongwook Bryce Kim
Language: English
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: Kwanjeong Educational Foundation
OpenAlex ID: W2773222027
