An Automated Quiet Sleep Detection Approach in Preterm Infants as a Gateway to Assess Brain Maturation

Anneleen Dereymaeker, Kirubin Pillay, Jan Vervisch, Sabine Van Huffel, Gunnar Naulaers, Katrien Jansen, Maarten De VosInt. J. Neural Syst 2017 · DOI

TL;DR

This paper introduces CLASS, a novel automated algorithm for detecting Quiet Sleep (QS) in preterm infants using EEG. This method offers a robust, age-independent approach to assess brain maturation, overcoming the limitations of time-consuming manual sleep staging.

What Problem Does This Paper Address?

Assessing brain maturation in preterm infants is crucial for identifying neurological risks and guiding interventions. Sleep state development, particularly Quiet Sleep (QS), is a key indicator of this maturation. However, manually labeling sleep stages from EEG is a highly specialized and time-intensive task, hindering large-scale studies and real-time monitoring. Existing automated methods are limited in their age range or robustness. This paper aims to bridge this gap by developing an automated, robust QS detection algorithm applicable across a wide range of postmenstrual ages (PMA) in preterm infants.

Methods

The core of this work is the CLuster-based Adaptive Sleep Staging (CLASS) algorithm. It operates in several stages:

  • Artifact Subspace Reconstruction (ASR): A novel artifact removal technique that uses PCA over sliding windows to identify and remove high-power, non-stationary artifacts from the EEG signal. This is a key improvement over traditional filtering methods.
  • Adaptive Segmentation (ASG): The cleaned EEG is downsampled and then segmented into variable-length segments based on changes in amplitude and frequency characteristics. This is more nuanced than uniform segmentation.
  • Feature Extraction and Clustering: Various time-domain and frequency-domain features are extracted from each segment. These features are then clustered using k-means to create "cluster-time profiles" that represent the evolving EEG characteristics.
  • QS Classification: QS is identified by analyzing fluctuations in the cluster-time profiles, as QS is characterized by higher EEG discontinuity. This involves de-trending, squaring, smoothing the profile, and applying a threshold to estimate QS periods.

Datasets

The study utilized a dataset of 89 EEG recordings from 26 preterm neonates, collected between the first and third week of life, with subsequent recordings every 2-3 weeks up to discharge. The recordings spanned a PMA range of 27 to 42 weeks.

  • Training Set: 34 recordings were used for algorithm development and optimization.
  • Test Set: 55 recordings were used for independent performance assessment.

The data was collected at the University Hospitals of Leuven, Belgium, and included neonates with strict inclusion criteria for normal neurodevelopmental outcome.

Key Insights and Ideas

  • QS as a primary marker: The paper highlights QS as a valuable initial marker for sleep assessment due to its relative discontinuity and lower artifact levels, making it suitable for automated analysis.
  • ASR's critical role: The introduction of Artifact Subspace Reconstruction (ASR) significantly improves the algorithm's robustness to artifacts, leading to better performance, particularly in reducing misclassification rates.
  • Adaptive Segmentation's advantage: ASG provides a more nuanced segmentation of the EEG signal compared to uniform segmentation, leading to improved clustering and classification.
  • Age-specific performance: CLASS demonstrates optimal performance in the 31-38 weeks PMA range, suggesting this is a critical period for QS development and assessment.
  • Maturational trends: The algorithm successfully captures maturational trends in QS characteristics, aligning well with visual assessments and providing stronger trends than non-state specific EEG analysis.

Key Results

  • Overall Performance: The CLASS algorithm achieved an excellent median Receiver Operating Characteristic (ROC) curve with an Area Under the Curve (AUC) of 0.9703 on the test set.
  • Performance in optimal age range (31-38 weeks PMA): Excellent Sensitivity (median 0.93-1.0), Detection Factor (DF) (median 1.0), Specificity (median 0.80-0.91), and low Misclassification Factor (MF) (median 0-0.25). MF was optimal (0) at 35-36 weeks PMA.
  • Impact of ASR: Comparing CLASS with ASR to CLASS without ASR (CLASS-noASR) showed significant improvements across all agreement measures, with CLASS-noASR having a higher MF (0.40) and lower AUC (0.85).
  • Comparison with other methods: CLASS outperformed the SAT% algorithm, which was found to be insufficient for robust QS detection.
  • Maturational Trends: QS-derived characteristics (Burst%, relative spectral powers) showed significant correlations with PMA, and these trends were more clearly defined and stronger when derived from QS periods compared to non-state specific EEG epochs.

Limitations and Future Work

  • Small Sample Size and Limited Recordings: The study acknowledges a relatively small sample size, especially for the youngest and oldest PMA groups, which can affect the robustness of findings at the extreme ages.
  • Ambiguity of Ground Truth: While inter-rater agreement was high (Kappa = 0.93), visual sleep classification remains somewhat ambiguous, relying on a combination of EEG and behavioral characteristics.
  • Performance at Extreme Ages: CLASS shows less certainty in classifying QS at <31 weeks PMA and higher misclassifications at >38 weeks PMA. This is attributed to the evolving nature of sleep states and the emergence of new QS patterns at these ages.
  • Future Work:
    • Automated AS Detection: The authors plan to extend algorithm development to include automated detection of Active Sleep (AS) alongside QS, to fully capture sleep-wake cycle dynamics.
  • Improved ASR: Exploring automatic age calibration for ASR could further enhance QS detection capabilities at extreme ages.
  • Larger Databases: Testing on new, well-described databases and incorporating input from multiple EEG experts could improve preterm and term neonatal EEG sleep interpretation.
  • Clinical Applications: Further exploration of the relationship between cerebral activity, brain development, and neurodevelopmental outcomes using CLASS for automated detection of abnormal preterm brain maturation.

Summary updated

BibTeX
@article{dereymaeker2017automated,
  author = {Anneleen Dereymaeker and Kirubin Pillay and Jan Vervisch and Sabine Van Huffel and Gunnar Naulaers and Katrien Jansen and Maarten De Vos},
  title = {{An Automated Quiet Sleep Detection Approach in Preterm Infants as a Gateway to Assess Brain Maturation}},
  journal = {Int. J. Neural Syst},
  year = {2017},
  doi = {10.1142/s012906571750023x},
}
Figure