TL;DR
This paper introduces an ontology-based decision support tool for automatically staging sleep stages using dual-channel EEG data. By integrating EEG features, sleep concepts, and contextual information into an "EEG-Sleep ontology" and employing a rule-based reasoning engine, the system achieves high accuracy in classifying sleep stages, offering a more objective and efficient alternative to manual staging.
What Problem Does This Paper Address?
Accurate sleep staging is crucial for diagnosing sleep-related disorders like insomnia and sleep apnea, which affect a significant portion of the population. Traditionally, this involves manual visual interpretation of polysomnography (PSG) data by experts, a process that is subjective, time-consuming, and prone to inter-observer inconsistencies. While automatic sleep staging methods exist, managing the vast amount of extracted EEG features and their contextual information remains a challenge. This paper aims to address this by proposing a structured, ontology-driven approach to represent and reason about EEG features for improved automatic sleep staging.
Methods
The proposed system is built around three core modules: an EEG-Sleep ontology, a rule set, and a reasoning engine.
- EEG-Sleep Ontology: This module acts as a knowledge representation framework. It encodes EEG features extracted from raw data, sleep-related concepts (like sleep stages and rules), and contextual information into a hierarchical structure. This ontology helps organize and manage the complex information related to sleep staging.
- Feature Extraction: A comprehensive set of EEG features is extracted, including linear methods (spectral power, Hjorth parameters), statistical methods (variance, skewness, kurtosis), and non-linear methods (spectral entropy, Shannon entropy, Kolmogorov entropy). A heuristic algorithm (sequential forward/backward selection) is used to select the most informative features.
- Rule Set Generation: A Random Forest (RF) algorithm is employed to train a rule set. This involves using the selected EEG features and the corresponding manually-labeled sleep stages (from visual interpretation) to generate IF-THEN rules that map feature patterns to specific sleep stages.
- Reasoning Engine: This module uses the EEG-Sleep ontology and the generated rule set to perform automatic sleep staging. It takes patient-specific EEG data, processes it through the ontology, and applies the rules to infer the sleep stage for each epoch.
Datasets
The study utilizes the Sleep-EDF database (EXPANDED), which contains 61 data recordings from 42 Caucasian subjects. The data includes dual-channel EEG from Fpz-Cz and Pz-Oz electrode locations, sampled at 100 Hz. The original sleep stages were defined according to R&K rules, and for this study, they were grouped into five states: Wake (WA), NREM1, NREM2, Slow Wave Sleep (SWS), and Rapid Eye Movement (REM).
Key Insights and Ideas
- Ontology for Structured Knowledge: Using an ontology provides a structured and hierarchical way to represent complex relationships between EEG features, sleep concepts, and contextual information, which is crucial for managing and reasoning about this data.
- Random Forest for Robust Rule Generation: The Random Forest algorithm proves effective for generating a robust rule set, handling noisy data, and providing good predictive performance, especially in multi-class problems like sleep staging.
- Dual-Channel EEG Sufficiency: The study demonstrates that dual-channel EEG data, when processed with an ontology-based approach, can be sufficient for achieving high accuracy in sleep staging, simplifying data acquisition compared to full PSG.
- NREM1 as a Challenge: The research highlights that NREM1 is the most challenging stage to classify accurately, likely due to its similar EEG patterns to WA, NREM2, and REM, and its transitional nature.
Key Results
The proposed ontology-based decision support tool achieved an average accuracy of 89.12% and a Kappa statistic of 0.81 on the test set. This performance was superior to other tested algorithms, including Bayesian Network (83.01%), C4.5 (78.31%), Multilayer Perceptron (86.42%), and Support Vector Machine (86.21%). The system also showed improved performance compared to other studies using subsets of the Sleep-EDF database. Specifically, the classification accuracies for WA, NREM2, SWS, and REM were high (81.92% to 93.43%), with NREM1 being the lowest at 57.86%.
Limitations and Future Work
- NREM1 Classification: The accuracy for NREM1 remains a limitation, and the authors suggest further research to improve its discrimination from other stages.
- Multimodal Fusion: While the current study focuses on dual-channel EEG, the authors propose future work to incorporate other physiological signals like EOG and EMG. They believe fusing these multimodal signals could further enhance recognition accuracy by complementing EEG data.
- Generalizability: While the study used a substantial dataset, further validation on larger and more diverse datasets would be beneficial to confirm the generalizability of the proposed approach.
Summary updated