A framework for multi-dimensional online temporal abstraction

Abstract

This dissertation details a framework for providing knowledge-based temporal reasoning and data analysis within a complex and high frequency data environment. The process of Temporal Abstraction (TA) involves a transformation from raw quantitative time-stamped data to a qualitative interval-based representation which can be matched directly to the domain expert's knowledge. Clinical data analysis involves a high degree of both contextual and temporal information. Diagnoses, prognosis and therapy incorporate knowledge of factors such as patient history, present medication and temporal factors like the characteristics of a particular condition with respect to time. Temporal abstraction provides a means to instill such properties into the data analysis process. Patient monitoring in an intensive care environment is an area where TA can be deployed to detect dangerous temporal trends or shifts in physiological data. The domain of the Neonatal Intensive Care Unit (NICU) is the context in which I demonstrate the proposals in this dissertation. The NICU data environment is both complex and multi-dimensional and represents a context where TA has not been previously applied. The process of Multidimensional Online Temporal Abstraction (MOTA) is the abstraction of physiological data across multiple data streams and data streamsets where a single patient gives rise to a single data stream-set and data is sourced from monitoring equipment connected to the patient. Applying TA to a multi-dimensional data environment provides a tool to support recent clinical research which proposes that aberrant behaviours across a number of physiological streams may have an additive effect and certain combinations thereof may be a sign of critical deterioration that could indicate a high likelihood of death or of brain injury in pre-term infants. Traditional patient monitoring equipment utilises the threshold alarming model where alerts are triggered upon the breaching of preset values. False alarms are common due to probe displacement or patient movement, temperature changes can adversely affect transducer accuracy and bad sensor positioning can result in low amplitude signals venturing outside prescribed ranges. Short lived excursions beyond thresholds often do not indicate clinically significant problems and could be dealt with by a more intelligent alarming model that incorporates a temporal dimension. Alarms can then be configured to sound if such excursions exist for longer than a pre-determined amount of time. Correlation between multiple physiological parameters is another desirable aspect of an improved alarming model. For example, the ability to simultaneously detect a decreasing trend in blood oxygen saturation with a decreasing trend in blood pressure would enable an early warning mechanism for a potentially dangerous situation. Cross correlation also offers the potential of reducing alarms through simultaneous measurement of the same parameter via different sources. Enabling the application of TA to data from multiple patients within the NICU also offers the potential for early detection of conditions such as Sepsis, Pneumothorax and Periventricular Leukomalacia (PVL) which have been shown to exhibit possible early warning characteristics before being diagnosable either through blood analysis in the case of sepsis, chest x-rays for Pneumothorax, or cranial sonography for PVL. TA is a well studied topic within the field of intelligent clinical data analysis but the research carried out to support this dissertation revealed that TA systems have difficulty coping with both high frequency and multi-dimensional data environments, typical of that found in neonatal intensive care. I present a software system called the MOTA framework which enables TA within such an environment and represents a step forward in the evolution of TA-based systems. The TA mechanisms of the MOTA framework are supported by a formal model of time, the Event Calculus, which I have adapted in order to facilitate online and incremental TA where abstractions are constructed on a sample by sample basis. The MOTA framework has undergone evaluation using clinically defined rules describing dangerous physiological patterns for infants in the NICU. These rules were executed on data captured from babies in the NICU, the results demonstrating that MOTA can be applied within a multi-dimensional and high frequency data environment such as Neonatal Intensive Care. Furthermore, potential is offered for improved patient monitoring leading to enhanced clinical management and better patient outcomes.

Date of Award	2009
Original language	English