Sleep Spindle Auto Detection Based on U-Net with Label Optimization Module

The automatic spindle detection algorithm based on deep learning technology requires high consistency of sleep spindle annotation, which crowdsourcing can significantly enhance through collective intelligence. To best of our knowledge, few studies have explored auto detection of sleep spindles based on crowdsourced datasets. MODA is the only crowdsourcing spindle annotation open access dataset until now, and this study utilized the MODA dataset to further enhance the automatic detection of sleep spindles using conventional neural network (CNN) and innovative label optimization module. To evaluate our model, we proposed and conducted two analysis methods including analysis by sample-event and by subject. Meanwhile, to validate our performance of our model, SUMO and EXP were considered as control and gold standard methods respectively; and various metrics related to by sample-event and by subject were compared among our model, SUMO and EXP. The results from analysis by sample-event showed that F1 score, Matthew's correlation coefficient (MCC), and Cohen's Kappa coefficient in our model were significantly higher than in SUMO under all overlap thresholds. Specifically, under a higher threshold of 0.6, F1, MCC and Kappa reached 0.82, 0.81, 0.81 in our model, 0.75, 0.74, 0.74 in SUMO, respectively. For analysis by subject, all spindle features extraction of our model were comparable to the EXP, including density (4.48 in our model vs 3.80 in EXP), duration (0.79 vs 0.84), amplitude (29.91 vs 30.09) and frequency (13.04 vs 12.92). However, the spindle density extracted by SUMO was significantly different from EXP (9.04 in SUMO vs 3.80 in EXP). In conclusion, our results suggested that the crowdsourced public dataset and label optimization module can further enhance the spindle detection capability of the convolutional neural network.