Automated sleep staging model for older adults based on CWT and deep learning.

Sleep staging plays a crucial role in the diagnosis and treatment of sleep disorders. Traditional sleep staging requires manual classification by professional technicians based on the characteristic features of each sleep stage. This process is time-consuming, and the accuracy of staging is heavily influenced by subjective factors. Currently, research on automatic sleep staging models based on deep learning has made significant progress. However, the automatic sleep staging models proposed by researchers seldom distinguish between age groups. With increasing age, changes in sleep architecture occur, and older adults experience a reduction in deep sleep duration. This age-related alteration makes older adults more susceptible to sleep disorders. Consequently, the automatic sleep staging problem for older adults is more challenging and warrants greater attention. This study first established a three-dimensional time‒frequency feature fusion map dataset based on continuous wavelet transform and determined the optimal channel signals from the Sleep-EDF expanded dataset. We subsequently proposed an automatic sleep model tailored for older adults, named RICM-SleepNet. This model employs Inception modules to extract features at multiple scales, uses the CBAM attention mechanism to further identify efficient features at different scales, and finally employs the multiscale connection structure to concatenate features from different stages, enhancing the model's feature utilization capability. RICM-SleepNet was subject to a performance evaluation on the three-dimensional time‒frequency feature fusion map dataset, yielding a sleep staging accuracy and κ value of 87.66% and 0.8492, respectively. Compared with the baseline models GoogLeNet, MobileNetV2, ShuffleNetV2, DenseNet121, RegNet, and ResNet50, RICM-SleepNet exhibited the highest recognition accuracy. To further validate the superiority of the RICM-SleepNet model, this study compared it with recent research methods that have demonstrated good performance in sleep staging. The results indicate that the proposed RICM-SleepNet model is superior to the other models in terms of performance. The Kruskal‒Willis test yielded a p value of 0.0014, indicating statistical significance. RICM-SleepNet attained the highest mean rank, underscoring its superior performance. In summary, the proposed multichannel automatic sleep staging model, RICM-SleepNet, shows promise in enhancing the accuracy and effectiveness of sleep staging, especially for older adults. Further validation and refinement of the model are warranted for its application in clinical settings and broader use in addressing sleep-related issues in the ageing population.