Abstract

Sleep spindles (SSs) appear in electroencephalogram (EEG) recordings during sleep stage N2, and they are usually detected through visual inspection by an expert. Labeling SSs in large datasets is time-consuming and depends on the expert criteria. In this work, we propose an unsupervised SS detector based on dictionary learning called the Unsupervised Sleep Spindle Detector (USSD). The proposed detector learns prototype SSs of different lengths (called atoms). An unsupervised adaptive threshold method based on the distribution of the automatically detected SS lengths is developed, which allows the adaptation of the USSD algorithm to different datasets in an unsupervised way. For each detection, the USSD estimates the probability of being an SS. The USSD performances on the labeled MASS-SS2 and INTA-UCH datasets yield F1-scores of $0.72 \pm 0.02$ and $0.72 \pm 0.04$ , respectively. The USSD outperforms the A7 and LUNA detectors, which are traditional unsupervised models. Next, we fine-tune the resulting USSD model with 20% of the labeled MASS-SS2 and INTA-UCH datasets, achieving F1 scores of $0.78 \pm 0.06$ and $0.75 \pm 0.05$ , respectively. In addition, the SSs detected by USSD on the unlabeled CAP dataset are used to pre-train a supervised deep learning method, which after fine-tuning with 20% of the MODA dataset, reaches an F1-score of $0.81 \pm 0.02$ .