Beyond Ground Truth in K-Complex Detection: A Waveform-Based SVM Classifier and the Limits of Expert Agreement
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Objective
K-complexes (KCs) are large-amplitude EEG events that represent N2 sleep stage and have been linked to sensory gating, sleep protection, and memory consolidation. Their detection remains limited by inter-rater variability in visual scoring and by the reliance of detectors on features that discard temporal information. We propose a two-stage detector that combines a rule-based candidate localization algorithm with a Support Vector Machine (SVM) classifier operating directly on the raw 2-seconds waveform, and we evaluate it against an adjudicated expert consensus of two different datasets.
Methods
Polysomnographic recordings from 10 healthy adults (Dataset 1) were independently annotated by two human scorers; discordant events were adjudicated by a senior expert, yielding 240 consensus KCs. The automatic classifier was evaluated using subject-level 10-fold Group K-Fold cross-validation and compared directly against the two human scorers under identical conditions. Cross-dataset generalization was further assessed on the public DREAMS database (Dataset 2) under both external and internal training criteria.
Results
The SVM classifier achieved the highest F1-score (79.4%) and accuracy (78.8%) among all scorers, with balanced recall (81.7%) and specificity (75.8%). Of the 58 false positives, 42 originated from events both experts had rejected yet displayed canonical KC morphology and received high classifier confidence (P(KC) > 0.7 in 45.2% of cases). This pattern was replicated on Dataset 2.
Conclusion
A waveform-based classifier matches expert performance and systematically flags morphologically valid KCs that fall outside conventional visual-scoring criteria.
Significance
These findings question the existence of an unambiguous ground truth for KC detection and support a data-driven redefinition of the event boundary, with implications for sleep staging and memory-consolidation research.
