AUTOMATED SLEEP STAGE DETECTION WITH A CLASSICAL AND A NEURAL LEARNING ALGORITHM – METHODOLOGICAL ASPECTS M. Schwaibold1, J. Schöchlin1, A. Bolz1,2 1
Medical Information Technology, FZI Forschungszentrum Informatik, Karlsruhe, Germany 2 Institute of Biomedical Engineering, University of Karlsruhe, Germany
[email protected]
Abstract For classification tasks in biosignal processing, several strategies and algorithms can be used. Knowledge-based systems allow prior knowledge about the decision process to be integrated, both by the developer and by self-learning capabilites. For the classification stages in a sleep stage detection framework, three inference strategies were compared regarding their specific strengths: a classical signal processing approach, artificial neural networks and neuro-fuzzy systems. Methodological aspects were assessed to attain optimum performance and maximum transparency for the user. Due to their effective and robust learning behavior, artificial neural networks could be recommended for pattern recognition, while neuro-fuzzy systems performed best for the processing of contextual information. Keywords sleep stage detection, biosignal processing, knowledge-based systems, artificial intelligence
Introduction Sleep stage scoring, as conceived by Rechtschaffen and Kales (R&K) [1], is complex, ambiguous, and it is difficult to devise and implement an algorithmic solution using classical digital signal processing methods. Nonetheless, it is the core task necessary for the evaluation of polysomno5 data channels EEG 1
EEG 2
EOG 1
EOG 2
EMG
graphy [2,3]. In R&K’s methodology, a human expert performs two major classification steps, as illustrated in Fig. 1: recognition of basic patterns in the 5 signal channels (2 EEG, 2 EOG and EMG) and scoring the current sleep stage based on the sequence of the detected sleep patterns. The sleep-related patterns, which have to be detected in the first step, can be distinguished as either periodic (alpha-, theta- and delta- activity) or transient (K-complexes, sleep spindles, vertex sharp waves, rapid / slow eye movements and body movements). A basic graphical illustration of common wave patterns is presented in [1], but such pattern recognition is primarily based on the expert’s experience and knowledge about typical waveforms. The second step consists of an evaluation of the rules defined by R&K. For both assessment steps, there is room for personal interpretation, and agreement among scoring experts can be as low as 50-70% for patients with Obstructive Sleep Apnea (OSA) [4].
Background Two algorithms for automated sleep stage recognition were assessed regarding their performance, with a focus on methodological and applicability aspects. Both algorithms
5 data channels
5 data channels
Preprocessing: 50 Hz - Filter, calibration to physical units
Preprocessing: 50 Hz - filter, calibration to physical units
Pattern recognition
Basic patterns
(based on waveform characteristics)
Feature extraction
EEG
Alpha
Theta
Delta
Vertex S.W. K-Complex
Artifact handling
Power Spectral Density Histogram of amplitudes Histogram of differences Waveform Signal power Correlation coefficients
Spindle
Pattern recognition
EOG REM
SEM
Tonus
Movement
Artificial neural networks (MLP, trained by RPROP)
EMG
Feature extraction
Sleep stage detection
Sleep stage detection
(Rule evaluation based on occuring patterns and their context)
Wake REM NREM 1 NREM 2 NREM 3 NREM 4 Movement
Fig. 1. Visual sleep stage scoring
Context layer Sequence of basic patterns in the relevant time window
Sleep stage recognition
Fig. 2. Conventional algorithm
Rule evaluation Neuro-fuzzy system (NEFCLASS)
Fig. 3. Processing steps of ARTISANA
showed very promising agreement rates with human scorers in the first validation trials. Detailed clinical validation studies are currently in progress. Classical signal processing algorithm: Classical automated sleep stage detection – as many biosignal processing approaches – consists of an artifact handling stage, feature extraction and a subsequent classification stage. This is illustrated in Fig. 2. The first 2 stages are based on threshold crossing checks for extracted parameters, like amplitude and frequency distributions or signal-template correlation coefficients. The classification stage combines its input information with logical rules to derive the sleep stage. This input consists of binary information about the incidence of the assessed features in the evaluated time window. Features could be considerable signal activity in certain frequency bands or waveforms similar to predefined templates. Knowledge-based algorithm (ARTISANA): The “Artificial Intelligence in Sleep Analysis” (ARTISANA) algorithm represents the detailed model of the described procedure as shown in Fig. 3. Components with neural learning capabilities were implemented to execute the two classification stages, pattern recognition and sleep stage detection. The design is intended to optimize both the quality and transparency of the calculated results. Examples of EEG, EOG and EMG sleep-related patterns are illustrated in Fig. 1. Waveforms differ between patients, these differences are also due to cross-talk, noise or combined mixed-frequency activity. These patterns are detected in 1 s intervals by Multi-Layer-Perceptron neural networks, which learned using the RPROP method. Neural networks adapt themselves to typical structures in their input vectors during a supervised learning procedure. The expert’s recognition experience is transferred to the automatic system with the accessible prior knowledge (frequency ranges, amplitude ranges and wave patterns) implemented into the input vector and the structure of the network. The sleep stage of a current interval not only depends on the sequence of patterns within the interval itself, but also on a certain time window. Thus, a context layer provides all pertinent information about preceding and succeeding patterns to the final decision stage, which is necessary to evaluate the rules of R&K. An example of contextual information is the time (in s) since the last rapid eye movement or K complex. The sleep stage detection stage consists of a hybrid neurofuzzy system, which mimics the definitions of R&K in predefined IF-THEN rules. During the supervised learning phase, the fuzzy membership functions in the premise parts of the rules are adapted. Finally, a defuzzification stage identifies the sleep stage for each 30 s interval using maximum search and summation of the seven membership grades of the 1 s intervals of the corresponding sleep stages.
Methodological aspects Two different types of self-learning signal processing systems were used for the two ARTISANA decision stages: artificial neural networks (ANNs) for pattern recognition and neural-fuzzy systems (NFS) for sleep stage detection. Characteristics and specific strengths were compared between these two knowledge-based systems and the classical signal processing approach (CSA). Development cost: First, information about the signal has to be extracted for the decision stage by digital signal processing algorithms, like power spectrum estimations or calculations of the frequency of occurrence of certain events. This stage is crucial for all signal processing approaches for reducing the information of the sampled signals to a small number of decisive parameters as input vectors to the decision stages. The developer has to apply a lot of gathered knowledge about the assessed process, and approved self-adapting algorithms for this task do not yet exist. Nevertheless, this stage is extremely important for the quality of the complete signal processing algorithm. Further procedures are different for all three signal processing approaches. In CSA, values of the input vector elements are typically compared to certain thresholds, and if the thresholds are exceeded, the data are utilized as binary information in classification rules. The focus of the development process is the definition of thresholds and inference rules. This is shown in Figure 4a. ANNs adapt themselves during a learning procedure to the typical values of the input parameters. The development does not directly focus on the signal processing task, but on the network topology, on the learning strategy and the compilation of a learning set, as can be seen in Figure 4b. Since efficient parameter optimization algorithms already exist, the development cost for neural network-based pattern recognition was the lowest. The specificity of NFS is the ability to combine both predefined rule definition and self-learning parameter optimization. This results in advantageous characteristics (see below), but also yields the highest development cost, as both rule definition and learning parameter configuration require expert knowledge, as shown in Figure 4c. Unfortunately, efficient parameter adaptation algorithms aren’t as available as those for neural networks. Transparency and optimization: Transparency of the inference engine is important for user acceptance, but also optimization strategies to be developed. While transparency can be guaranteed in CSA, optimization possibilities make parameter extraction and the inference stage more complex. In ANN, the optimization process is partially automated. If neural learning does not lead to a satisfactory result, extracted features or the composition of the learning set can be modified. The effects of both optimization strategies can hardly be predicted and require intensive experience. This obstacle is closely related to the black-box character of ANN, resulting in an extremely limited transparency. NFS provide the largest flexibility of parameter optimization and systematic strategies, due to a reasonably high transparency. Fuzzy information processing allows pa-
a)
CSA feature extraction input vector
rule definition
parameter adjustment
b)
ANN feature extraction input vector
rule definition
learning procedure
c)
parameter adjustment
learning set
NFS feature extraction input vector
rule definition
learning procedure
parameter adjustment
learning set
high low influence capabilities for optimization associated with required development efforts
Fig. 4: Development expenditures and optimization capabilities of CSA, ANN and NFS rameters to be continuously adapted. Since there are many different degrees of freedom, optimization of NFS is the most complex procedure yielding the highest potential. Prior knowledge: The implementation of prior knowledge proved to be a key factor for any recognition or detection algorithm. Due to the mentioned black-box character of ANNs, prior knowledge about the relationship of input parameters can hardly be implemented, even if available. In NFS, prior knowledge about parameter relations can be implemented directly and effectively.
Discussion All three strategies for classification tasks in biosignal processing revealed specific strengths and weaknesses for the decision tasks required by a sleep stage recognition framework. CSA proved to be an efficient strategy to achieve a relatively robust algorithm with reasonable development costs. Since there is no way to automatically adapt parameters, optimization is complicated, typically involving an ex-
tremely limited number of input parameters and manual adaptation of the complete set of degrees of freedom. The assessed conventional algorithm was designed according to this strategy, resulting in the robust detection of typical signal intervals, but in increased difficulties for complex decision cases. ANNs showed robust learning capabilities even with limited prior knowledge. Thus, they are specifically suited for complex decision tasks with a large number of input parameters and limited knowledge about detailed relations. This is the case for pattern recognition in EEG, EOG and EMG when extracting a larger set of parameters from the signal. ANNs learn from given samples, much like a human scorer does from a guiding expert. The compilation of learning sets should be performed carefully and with great amount of experience. NFS as a combination of a rulebased inference engine and a neurally learning network shows great potential for biosignal classification. Their transparency can support the user and developer to come to the same decision as the automated system. Nonetheless, development expenditures are high and detailed knowledge about the decision process is required. Thus, NFS were selected to process contextual information about occurred patterns where the definitions of Rechtschaffen and Kales could directly be transferred into a rulebase. By combining ANN and NFS, a modular sleep stage detection algorithm ARTISANA was designed, providing maximum transparency and flexible extension possibilities, e.g. to assess the sleep microstructure or to visualize intermediate results to the user. A more efficient learning algorithm for NFS has been developed and is being further improved.
Acknowledgement The author thanks MCC GmbH&Co. KG for their assistance and support.
References [1] A. Rechtschaffen, A. Kales, A manual of standardized terminology, techniques and scoring system for sleep stages of human subjects, U.S. Department of Health, National Institutes of Health Publication No. 204, 1968. [2] K.E. Bloch, “Polysomnography: a systematic review”, Technol. And Health Care, vol. 5, pp. 285-305, 1997. [3] T. Penzel, R. Conradt, “Computer based sleep recording and analysis. Review article”, Sleep Med. Rev., vol. 4, pp. 131-148, 2000. [4] R.G. Norman, I. Pal, C. Stewart, J.A. Walsleben, D. M. Rapoport, “Interobserver agreement among sleep scorers from different centers in a large dataset“, Sleep, vol. 23-7, pp. 901-908, 2000. [5] M. Schwaibold, B. Schöller, T. Penzel, A. Bolz, “Artificial Intelligence in Sleep Analysis (ARTISANA) – Modellierung des visuellen Vorgehens bei der Schlafklassifikation“, Biomedizinische Technik, vol. 46-5, pp. 129-132, 2001. [6] Unpublished data.
