Towards symbolization using data-driven extraction of local trends for ICU monitoring.

Laboratoire d'Automatique et Informatique Industrielle de Lille, Bâtiment P2 Cité Scientifique, Villeneuve d'Ascq, France.
Artificial Intelligence in Medicine (Impact Factor: 1.36). 08/2000; 19(3):203-23. DOI: 10.1016/S0933-3657(00)00046-4
Source: PubMed

ABSTRACT We propose a methodology for the extraction of local trends from a stream of data. It has been designed to suit the needs of interpretation-oriented visualization and symbolization from ICU monitoring data. After giving implementation details for efficient computation of local trends, we propose the use of a characteristic analysis span for each variable. This characteristic span is obtained from a set of criteria that we compare and evaluate in regard of analysis of ICU monitoring data gathered within the Aiddaig project. The processing results in a rich visual representation and a framework for the local symbolization of the data stream based on its dynamics.

  • Source
    [Show abstract] [Hide abstract]
    ABSTRACT: Intelligent clinical data analysis systems require precise qualitative descriptions of data to enable effective and context sensitive interpretation to take place. Temporal abstraction (TA) provides the means to achieve such descriptions, which can then be used as input to a reasoning engine where they are evaluated against a knowledge base to arrive at possible clinical hypotheses. This paper surveys previous research into the development of intelligent clinical data analysis systems that incorporate TA mechanisms and presents research synergies and trends across the research reviewed, especially those associated with the multi-dimensional nature of real-time patient data streams. The motivation for this survey is case study based research into the development of an intelligent real-time, high-frequency patient monitoring system to provide detection of temporal patterns within multiple patient data streams. The survey was based on factors that are of importance to broaden research into temporal abstraction and on characteristics we believe will assume an increasing level of importance for future clinical IDA systems. These factors were: aspects of the data that is abstracted such as source domain and sample frequency, complexity available within abstracted patterns, dimensionality of the TA and data environment and the knowledge and reasoning underpinning TA processes. It is evident from the review that for intelligent clinical data analysis systems to progress into the future where clinical environments are becoming increasingly data-intensive, the ability for managing multi-dimensional aspects of data at high observation and sample frequencies must be provided. Also, the detection of complex patterns within patient data requires higher levels of TA than are presently available. The conflicting matters of computational tractability and temporal reasoning within a real-time environment present a non-trivial problem for investigation in regard to these matters. Finally, to be able to fully exploit the value of learning new knowledge from stored clinical data through data mining and enable its application to data abstraction, the fusion of data mining and TA processes becomes a necessity.
    Artificial Intelligence in Medicine 02/2007; 39(1):1-24. · 1.36 Impact Factor
  • [Show abstract] [Hide abstract]
    ABSTRACT: The trend patterns of vital signs provide significant insight into the interpretation of intraoperative physiological measurements. We have modeled the trend signal of a vital sign parameter as a generalized hidden Markov model (also known as a hidden semi-Markov model). This model treats a time series as a sequence of predefined patterns and describes the transition between these patterns as a first-order Markov process and the intra-segmental variations as different dynamic linear systems. Based on this model, a switching Kalman smoother combines a Bayesian inference process with a fixed-point Kalman smoother in order to estimate the unconditional true signal values and generates the probability of occurrence for each pattern online. The probabilities of pattern transitions are tested against a threshold to detect change points. A second-order generalized pseudo-Bayesian algorithm is used to summarize the state propagation over time and reduces the computational overhead. The memory complexity is reduced using linked tables. The algorithm was tested on 30 simulated signals and 10 non-invasive-mean-blood-pressure trend signals collected at a local hospital. In the simulated test, the algorithm achieved a high accuracy of signal estimation and pattern recognition. In the test on clinical data, the change directions of 45 trend segments, out of the 54 segments annotated by an expert, were correctly detected with the best performing threshold, and with the introduction of only 8 false-positive detections. The proposed method can detect the changes of trend patterns in a time series online, while generating quantitative evaluation of the significance of detection. This method is promising for physiological monitoring as the method not only generates early alerts, but also summarizes the temporal contextual information for a high-level decision support system. Copyright © 2009 John Wiley & Sons, Ltd.
    International Journal of Adaptive Control and Signal Processing 06/2009; 24(5):363 - 381. · 1.22 Impact Factor
  • [Show abstract] [Hide abstract]
    ABSTRACT: This paper presents an alarm validation system dedicated to patient monitoring in intensive care units (ICU). Several physiological signals are continuously acquired and an on-line trend extraction method is implemented for each one. A trend is a succession of contiguous semi-quantitative episodes, expressing the time evolution of a signal with several symbols. The difference between the trend and the signal is considered as a residual. In this paper, trend extraction is based on several thresholds that are adapted on-line, following the signal variations. Multivariate change indices are further deduced from the trends and the residuals. They provide an indication of changes to patient hemodynamic and respiratory state. An alarm validation system based on these indices is then proposed, which uses fuzzy decision making. Whenever a monitoring system sets off an alarm, the system proposed carries out a backward analysis of the physiological variables monitored. The system enables various policies to be implemented: filtering of false alarms due to artifacts, confirmation of true alarms due to a patient state change. The system was tested on more than 50 h of data recorded on adult patients in an ICU unit, when 105SpO2 alarms were set off by a fixed threshold alarm system. The comparison between the decision made on-line by the validation system and the decision made by a medical expert for each of these alarms showed that the system is able to recognize 100% of true alarms and filter 50–80% of false alarms. Copyright © 2009 John Wiley & Sons, Ltd.
    International Journal of Adaptive Control and Signal Processing 06/2009; 24(5):382 - 408. · 1.22 Impact Factor