A comparative study of algorithms for sleep spindle detection

The automated detection of sleep spindles constitutes a central methodological and computational challenge in the study of sleep and its electroencephalographic (EEG) correlates. From the perspective of the research workflow, the precise extraction of these transient microstructures represents a preliminary, foundational processing step. Nevertheless, it establishes the necessary paradigm for properly evaluating spindles from a clinical and neurophysiological standpoint, enabling researchers to investigate their temporal dynamics, individual traits, and functional role. Formally, a sleep spindle is a transient burst of neural oscillatory activity generated within the thalamic and corticothalamic networks. It occurs within the sigma frequency band, typically oscillating between 11 and 16 Hz, with a minimum continuous duration of 0.5 seconds and rarely exceeding 2 to 3 seconds. Serving as a defining electrophysiological feature of non-rapid eye movement (NREM) sleep, sleep spindles are predominantly observed during stage N2, but continue to occur at a reduced frequency throughout deep slow-wave sleep (N3). To date, the clinical gold standard for the detection and characterization of sleep spindles strictly relies on visual scoring by expert neurophysiologists—a manual paradigm burdened by severe methodological limitations. This manual procedure is extremely time-consuming and largely inadequate for processing modern large-scale polysomnographic datasets. Furthermore, it is intrinsically affected by high intra- and inter-scorer variability. This fundamental subjectivity, leading to poor reproducibility of results, has served as the primary drive for the extensive development and proliferation of numerous automated detection algorithms. In this thesis, four specific automated detection algorithms are analyzed, structurally implemented, and operationalized: A7, YASA, ConceFT, and SUMO. The first three belong to the family of parametric and deterministic methods. Specifically, A7 and YASA operate using moving threshold logics based on relative sigma power and correlation in the time domain, while ConceFT employs an advanced synchrosqueezing time-frequency transform to reallocate spectral energy. The fourth algorithm, SUMO, represents a machine learning topology based on deep convolutional neural networks. These frameworks were meticulously adapted, optimized, and tested on electroencephalographic data from the SS2 cohort of the Montreal Archive of Sleep Studies (MASS) dataset, and their detection performances were statistically evaluated against human annotations. A central focus of this research was the in-depth investigation of the specific factors causing structural failures or highly unstable performances in these automated frameworks. The comparative analysis revealed that algorithmic efficacy is not uniquely determined by underlying mathematical complexity. Instead, detection success heavily depends on the intrinsic physical properties of the biological signal, particularly the signal-to-noise ratio (SNR) and phase non-stationarity driven by macro-artifacts. Moreover, the human ground-truth itself emerged as a critical variable: algorithmic performance metrics collapse or peak drastically depending on whether the reference expert adopts a highly restrictive or inclusive scoring topology, or exhibits annotation fatigue. To overcome these physical and methodological limitations, an original pre-analytical "triage" pipeline for dynamic algorithm routing was formalized. By computing a deterministic Signal Quality Index (SQI) directly within the N2 sleep epochs, the pipeline maps the physical properties of the data, routing stable, high-SNR signals to the precision-oriented deep learning filters of SUMO, while diverting artifact-heavy EEGs toward the intrinsic spectral robustness of ConceFT. However, the introduction of a second human expert into the evaluation matrix revealed an unexpected paradigm shift. Ultimately, the findings of this thesis challenge the pursuit of a single "flawless" algorithm, demonstrating that the choice between a dynamic hybrid routing system and a robust universal spectral model fundamentally depends on which human cognitive style the machine is tasked to replicate.