Yamabe et al. - 2019 - MC-SleepNet Large-scale Sleep Stage Scoring in Mice by Deep Neural Networks

[gsaurabhr]

Read and summarize key findings in this thread.

[Tanvig]

Key points from the paper:

• MC-SleepNet is an automated method designed to identify the sleep stage of each epoch in a time series of EEG and EMG signals in mice. It consists of signal preprocessing, feature extraction using CNN, and scoring using bi-LSTM. It uses both EEG and EMG signals.

• The CNN is able to extract effective features from EEG and EMG signals. It can also automatically identify features related to individual differences and noise, thus improving scoring accuracy and robustness.

• The LSTM network is able to handle long time-series data, which is essential for modeling sleep cycles and transitions. It includes a "forget gate" mechanism to manage the retention and disposal of past information, making it suitable for long sequences. For bidirectionality, two LSTM layers are utilized—one processes data in chronological order, while the other processes it in reverse. This dual approach allows consideration of both future and past states, enhancing the model's ability to understand sleep stage transitions.

• For signal preprocessing, the EMG signals were preprocessed using a moving root mean squared filter with a 1-second width to emphasize amplitude (for identifying sleep stages). The EEG signals were analyzed directly without preprocessing to retain both frequency-domain and time-domain features. The signals are divided into 20-second epochs for subsequent analysis.

• For feature extraction, CNN is used to locate effective features from the EEG and EMG signals automatically. The EEG signals were processed with two CNN blocks featuring wide and narrow filters to capture low and high-frequency features. The EMG signals were processed with one CNN block featuring wide filters, given the lack of high-frequency components due to preprocessing.

• The scoring module consists of a bi-LSTM block, a fully connected (FC) block, and a softmax layer, modeling the relationship between features and sleep stages while considering sleep stage transition rules. The bi-LSTM block processes 25 consecutive epochs to model sleep stage transitions. The FC Block focuses on the target epoch to detect "isolated" epochs with differing stages from neighbouring epochs.

• MC-SleepNet training involves two steps: pretraining and fine-tuning. In pretraining, the feature extraction module (CNN) was optimized by temporarily replacing the scoring module with a softmax layer and using categorical cross-entropy loss. In fine-tuning, the softmax layer is replaced with the original scoring module, and then the entire system is retrained, and parameters are fine-tuned using a lower learning rate to prevent overfitting.

• The evaluation of MC-SleepNet is conducted on a large-scale dataset comprising 4,200 biological signal records from mice. The model achieved a scoring accuracy of 96.6%. For noisy records, the accuracy was 95%.

• With noisy data, the model tended to output non-REM stage labels more frequently. This results in a reduction in the recall for WAKE and REM stages and a decrease in the precision of non-REM stage labelling. A rescoring model was then developed to address the lower recall for the REM stage. It re-evaluates epochs with non-REM labels and low certainty values.