A method for automated temporal knowledge acquisition applied to sleep-related breathing disorders

https://doi.org/10.1016/S0933-3657(01)00089-6Get rights and content

Abstract

This paper presents a method for the discovery of temporal patterns in multivariate time series and their conversion into a linguistic knowledge representation applied to sleep-related breathing disorders. The main idea lies in introducing several abstraction levels that allow a step-wise identification of temporal patterns. Self-organizing neural networks are used to discover elementary patterns in the time series. Machine learning (ML) algorithms use the results of the neural networks to automatically generate a rule-based description. At the next levels, temporal grammatical rules are inferred. This method covers one of the main “bottlenecks” in the design of knowledge-based systems, namely, the knowledge acquisition problem. An evaluation of the rules lead to an overall sensitivity of 0.762, and a specificity of 0.758.

Introduction

Due to the increased performance and storage capacities of computers in the past years, a huge amount of data has been measured in different application domains, such as industrial processes and medical applications. Often continuous signal data are acquired in order to enhance the diagnosis or control of the underlying process. These signal data are frequently recorded for time periods resulting in multi-signal recordings that describe the underlying process.

For example, consider a human being that suffers from daytime sleepiness. Sleepiness is often caused by insufficient sleep, which may result from shift work or poor sleep hygiene. However, the most common medical disorder causing excessive daytime sleepiness appears to be sleep apnoea [39]. For the diagnosis of sleep apnoea the dynamics of physiological parameters, such as respiration and heart rate, must be recorded and evaluated [2], [33]. Moreover, for an accurate analysis of sleep apnoea, a large number of parameters must be evaluated. These are sleep-related signals (EEG, EOG, EMG), signals concerning respiration (airflow, ribcage and abdominal movements, oxygen saturation, snoring), and circulation-related signals (ECG, blood pressure) [34], [40]. The continuous recording of these signals is known as a polysomnography (PSG) [30]. The identification of the different types of sleep apnoea, namely, apnoea and hypopnoea, can be carried out just using signals directly related to respiration [2], [30]. Counting the number of individual apnoeas and hypopnoeas/h of sleep, which is the apnoea-index and the hypopnoea-index, gives a measure for the extension of the disorder. The respiratory disturbance index (RDI) is the sum of apnoeas and hypopnoeas /h of sleep. It can be seen as pathological, when the RDI is larger than 20 events/h. After 40 events/h, the patient has to be referred to therapy.

The diagnosis of sleep apnoea, more exactly sleep-related breathing disorders (SRBDs), is made by physicians with the help of technical assistants who perform a visual classification of the different types of apnoeas using a PSG. An automatic identification of the SRBDs is a very hard task, since a great number of signals must be analyzed simultaneously [41]. Even for the same patient, quite different patterns may occur for the same SRBD type. In addition, a strong variation of the duration of each SRBD may occur [31].

In order to solve this kind of problems, we propose a step-wise identification of SRBDs. At different abstraction levels, introduced by the method, a progressively higher temporal-abstraction of the underlying signals may be achieved. Therefore, methods from artificial intelligence (AI), artificial neural networks (ANNs), computer science and statistics are used. The main advantage of ANNs lies in their effectiveness with regard to applications with sub-symbolic data, such as noisy or even inconsistent real-world data. Such a sub-symbolic processing appears to be more adequate for pattern recognition problems [7]. Usually, ANNs with supervised learning are used to solve these problems. These ANNs, however, have some limitations when dealing with unclassified data, as it frequently occurs in large databases. In order to overcome this disadvantage of supervised learning methods, we propose to use self-organizing neural networks, as they may learn structures from high dimensional data without any prior information, i.e. a pre-classification of the patterns [8].

The generation of a temporal knowledge representation for the discovered patterns is another main aim of our approach. This requires the abstraction of the time series into higher-level concepts meaningful for a specific domain. This task is often referred to as temporal-abstraction, and must include an explicit representation of time. A well-known approach in medicine is the RÉSUMÉ knowledge-based temporal-abstraction system, envisaged as a problem-solving method [38]. During runtime certain types of domain-specific clinical knowledge are assumed, represented for instance as classifications rules or temporal-semantic properties. This approach, however, assumes a language that permits knowledge acquisition from domain experts [12]. Our method, however, automatically generates a temporal-abstraction for the multivariate time series in the form of temporal grammatical rules intelligible for human beings, such as medical experts. This includes the use of machine learning (ML) algorithms. An implementation of these rules in Prolog allows the use of traditional AI-technologies, such as knowledge-based systems, that have been successful in several application domains, namely diagnosis, control, and planning [21].

The advantages of both technologies, AI-technologies versus ANN-technologies, are wide-ranging. However, both approaches show up some limitations, for instance, the incapacity of ANN to explain their behavior and the problem of knowledge acquisition for AI-systems. Recently, there has been an increased interest in hybrid systems that integrate both technologies [15]. Usually, hybrid systems entail several modules that co-operate with one another. Each module is then implemented in a different technology. We are mainly interested in hybrid systems that perform a knowledge conversion, i.e. a transition between different knowledge representation forms, i.e. a sub-symbolic and a symbolic knowledge representation [44]. Previous applications that perform a knowledge conversion do not consider temporal dependencies among the data [43], [46]. This paper, however, presents a method that enables a temporal knowledge conversion [16], [18] for multivariate time series, applied to sleep apnoea.

In Section 2, we present some basic concepts of special self-organizing neural networks for the discovery of patterns in high-dimensional data. The method for temporal knowledge conversion is introduced in Section 3. Section 4 includes a detailed description of the experiments made with data from three patients with SRBDs. The evaluation of the method applied to SRBDs will be given in Section 5. Section 6 includes a discussion of the proposed method and its application to SRBDs. Conclusions are presented in Section 7.

Section snippets

Self-organizing neural networks for the discovery of patterns in high-dimensional data

ANNs may be classified according to their learning principles into ANNs with supervised and ANNs with unsupervised learning. The most popular supervised learning algorithm is the back-propagation algorithm [37]. This type of ANN is often used for pattern recognition problems, such as recognition of images representing hand-written characters [7]. This requires, however, a previous classification of the data, for example, from a domain expert.

We are mainly interested in ANNs with unsupervised

A method for automated temporal knowledge acquisition

This section presents the recently developed method for temporal knowledge conversion (TCon) [16], [18] that performs an automated temporal knowledge acquisition (see Fig. 1). The following constraints should be considered:

  • The method should be simple enough, in order to enable the discovery of inherent patterns in multivariate time series using clustering methods, such as ANNs with unsupervised learning, as well as an automated generation of a rule-based description of those patterns using

The method applied to sleep-related breathing disorders

The method was applied to a sleep disorder with a high prevalence, called sleep apnoea. More exactly, SRBDs consist of various types among which sleep apnoea is best known [2], [33]. For an analysis of sleep apnoea, a large number of parameters must be evaluated, namely sleep-related signals (EEG, EOG, EMG), signals concerning respiration (airflow, ribcage and abdominal movements, oxygen saturation, snoring), and circulation related signals (ECG, blood pressure) [40]. The parallel recording of

Results

For the evaluation of the method applied to SRBDs, a qualitative evaluation of the knowledge base [4] was made. A questionnaire was developed for carrying out a structured interview of the domain expert that examines the following points: identification of all types of disorders related to temporal patterns, sequences and events according to domain knowledge, valid duration of temporal patterns and sequences, interpretability and intelligibility of TG-rules at different levels, naming of

Discussion

The method presented here has two main aims, namely temporal pattern discovery with unsupervised methods and temporal knowledge extraction with rule induction algorithms.

For the discovery of unknown temporal patterns in time series several approaches have already been proposed. For instance, dynamic time warping uses a dynamic programming approach to align the time series [5]. Since the time axis is stretched (or compressed) to achieve a reasonable fit, a template may match a variety of actual

Conclusion

This paper presents an application to SRBDs of a method for temporal knowledge conversion [18]. The main idea lies in the introduction of several abstraction levels to attain a segmentation of this highly complex problem into several subtasks. The method has two main goals, namely temporal pattern discovery and temporal knowledge extraction. Since it may be used as a tool for automated temporal knowledge acquisition, it covers one of the main “bottlenecks” in the design of knowledge-based

References (47)

  • Bock HH. Clustering and neural networks. In: Rizzi A, et al. (Eds.), Advances in Data Science and Classification,...
  • Bontempi G, Birattari M, Bersini H. Local learning for iterated time-series prediction. In: Bratko I, Dzeroski S,...
  • Bukharaev RG. Theorie der stochastischen Automaten. Stuttgart: Teubner Verlag,...
  • Bürsner S, Spreckelsen C, Nebe J, Spitzer K. Epistemologische Adäquatheit bei der Akquisition und formalen...
  • Chakravarty S, Shahar Y. Acquisition and analysis of periodic patterns in time-oriented clinical data. Technical report...
  • Cimino JJ, Barnett GO. Automatic knowledge acquisition from MEDLINE. Methods of information in medicine. Schattauer...
  • Fukunaga K. Introduction to statistical pattern recognition. 2nd ed. Boston: Academic Press,...
  • Goonatilake S, Khebal S. Intelligent hybrid systems. New York: Wiley,...
  • Guimarães G. Eine Methode zur Entdeckung von komplexen Mustern in Zeitreihen mit Neuronalen Netzen and deren...
  • Guimarães G, Ultsch A. A symbolic description for patterns using definitive clause grammars. In: Klar R, Opitz O,...
  • Guimarães G, Ultsch A. A method for temporal knowledge conversion. In: Hand DJ, Kok JN, Berthold MR, editors. Advances...
  • Guralnik V, Srivastava J. Event detection from time series data. In: Proceedings of the Fifth ACM SIGKDD International...
  • Kaski S, Kohonen T. Exploratory data analysis by self-organizing map: structures of welfare and poverty in the world....
  • Cited by (21)

    • A review on time series data mining

      2011, Engineering Applications of Artificial Intelligence
      Citation Excerpt :

      These descriptions are based on short and long-term trends, which are detected using the wavelet transform. Guimaraes and Ultsch (1999) propose an approach to transit patterns in multivariate time series to a linguistic description. Different abstraction levels’ temporal grammatical rules are extracted from the results of neural networks and other exploratory methods.

    • Symbolic representation based on trend features for knowledge discovery in long time series

      2015, Frontiers of Information Technology and Electronic Engineering
    • Comparison of neural network back propagation algorithms for early detection of sleep disorders

      2015, Conference Proceeding - 2015 International Conference on Advances in Computer Engineering and Applications, ICACEA 2015
    • Soft computing technique based on ANFIS for the early detection of sleep disorders

      2015, Conference Proceeding - 2015 International Conference on Advances in Computer Engineering and Applications, ICACEA 2015
    • Change detection with Kalman filter and CUSUM

      2010, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
    View all citing articles on Scopus
    1

    Tel.: +49-6421-2866435; fax: +49-6421-2865450.

    2

    Tel.: +49-6421-2822185; fax: +49-6421-2828902.

    View full text