End-to-End Deep Learning Model For Automatic Sleep Staging Using Raw PSG Waveforms

Deep learning has seen significant progress over the last few years, especially in computer vision, where competitions such as the ImageNet challenge have been the driving factor behind many new model architectures far superior to humans in image recognition. We propose a novel method for automatic sleep staging, which relies on current advances in computer vision models eliminating the need for feature engineering or other transformations of input data.

By exploiting the high capacity for complex learning in a state of the art object recognition model, we can effectively use raw PSG signals to detect and classify sleep stages in a robust and reliable way. Methods: A total of 2322 PSG studies from the Wisconsin Sleep Cohort were used in this study. Central and occipital EEG, left and right EOG, and chin EMG signals were extracted from all PSGs and subjected to initial pre-processing of zero-phase Butterworth bandpass filters with AASM-specified cutoffs.

The raw signals were then segmented into 30 s epochs and fed as inputs to a novel deep neural network model based on the ResNet-50 architecture. The model was optimized over cross-entropy loss with respect to annotated scorings using the Adam optimizing algorithm and trained on a subset of 1858 PSGs.

Hyperparameters were tuned using 40 iterations of random search in relevant hyperparameter intervals. Best performing model was selected based on performance measured by overall accuracy on a hold-out validation set of 232 PSGs. Results: Training accuracy, precision and recall were 84.93%, 97.42% and 97.02%, respectively.

Evaluating on the validation set yielded an overall accuracy of 85.07% and overall precision/recall of 98.54% and 95.72%, respectively. Conclusion: Preliminary results indicate that state of the art deep learning models can effectively be used to classify sleep stages using untransformed PSG signals.

We will perform further testing on independent datasets to enhance the model’s utility.