Abstract
Introduction
Manually scoring polysomnograms is both time-consuming and labor-intensive. It also increases variability in care. Generating features for use as components within larger models is an important part of building highly accurate auto-scoring systems. In this study, we examined the use of time-frequency data representations in combination with a convolutional neural networks (CNN).
Methods
We used just six (6) pre-scored polysomnograms from the MrOS dataset in this analysis. Only one electroencephalography (EEG) and one electrooculography (EOG) channel were extracted from each polysomnogram and split into 30 second epochs. Visual representations of each epoch in the time-frequency domain were generated using Morlet wavelets, then divided into training and validation sets in a 4:5 distribution. We then re-trained a ResNet-50 CNN using transfer learning to classify sleep stage based on the time-frequency representations.
Results
A total of 4971 epochs were generated. Of those, 1242 epochs formed the validation set. Performance was high for identifying Stage W with an accuracy of 94.2% (295/313 epochs). However, performance for other stages was considerably lower. Stage N3 was predicted correctly in 68.0% of cases (138/203 epochs), although in 60/75 cases of misclassification the predicted class was Stage N2. Similarly, Stage N2 was predicted correctly in 62.0% of cases (183/295 epochs), and in 63/112 cases of misclassification the predicted class was Stage N3. Accuracy for Stage REM was 64.9%. Stage N1 prediction was poor (22.0% accuracy), likely due to insufficient representation in the sample (< 10% of epochs).
Conclusion
This exploratory analysis of the use of time-frequency representations in conjunction with a CNN demonstrates some promise, especially with respect to prediction of Stage W using this technique. Inclusion of additional data channels and larger sample size would likely improve accuracy.
Support
RS - ASPIRE Fellowship (sponsored by the American Thoracic Association).