Temporal pattern recognition in multiparameter ICU data

Abstract

Intensive Care Unit (ICU) patients are physiologically fragile and require vigilant monitoring and support. The myriad of data gathered from biosensors and clinical information systems has created a challenge for clinicians to assimilate and interpret such large volumes of data. Physiologic measurements in the ICU are inherently noisy, multidimensional, and can readily fluctuate in response to therapeutic interventions as well as evolving pathophysiologic states. ICU patient monitoring systems may potentially improve the efficiency, accuracy and timeliness of clinical decision-making in intensive care. However, the aforementioned characteristics of ICU data can pose a significant signal processing and pattern recognition challenge---often leading to false and clinically irrelevant alarms. We have developed a temporal database of several thousand ICU patient records to facilitate research in advanced monitoring systems. The MIMIC-II database includes high-resolution physiologic waveforms such as ECG, blood pressures waveforms, vital sign trends, laboratory data, fluid balance, therapy profiles, and clinical progress notes over each patient's ICU stay. We quantitatively and qualitatively characterize the MIMIC-II database and include examples of clinical studies that can be supported by its unique attributes. We also introduce a novel algorithm for identifying "similar" temporal patterns that may illuminate hidden information in physiologic time series. The discovery of multi-parameter temporal patterns that are predictive of physiologic instability may aid clinicians in optimizing care. In this thesis, we introduce a novel temporal similarity metric based on a transformation of time series data into an intuitive symbolic representation.

(cont.) The symbolic transform is based on a wavelet decomposition to characterize time series dynamics at multiple time scales. The symbolic transformation allows us to utilize classical information retrieval algorithms based on a vector-space model. Our algorithm is capable of assessing the similarity between multi-dimensional time series and is computationally efficient. We utilized our algorithm to identify similar physiologic patterns in hemodynamic time series from ICU patients. The results of this thesis demonstrate that statistical similarities between different patient time series may have meaningful physiologic interpretations in the detection of impending hemodynamic deterioration. Thus, our framework may be of potential use in clinical decision-support systems. As a generalized time series similarity metric, the algorithms that are described have applications in several other domains as well.