# EEG Papers

### Deep Sleep Net(2017)
>[Notes](https://hackmd.io/cW_57wdUQ7-2dVyblDbrnQ)
- Inputs raw single channel EEG(F4-EOG left,Fpz-Cz & Pz-Oz).
- Uses CNN and bidirectional LSTMs and independent splitting.
- Sleep edf accuracy,MF1 scores are 82, 76.9 using Fpz-Cz and 79.8, 73.1 using Pz-Oz.
- F1 score for N1 is 46.6.
- MASS accuracy,MF1 scores are 86.2, 81.7 and F1 score for N1 is 59.8(Uses F4-EOG)

### Deep Conv Net for EEG anaysis(2017)
>[Notes](https://hackmd.io/lNFWPqujQcWnH4oVMf-hig)
- Uses multitaper spectral analysis to preprocess EEG.
- Uses a VGG model along with transfer learning(using ILSVRC data) and independent splitting.
- Comparison metrics with existing networks not clearly displayed.
- VGG FE has F1 score of 58 on N1 and VGG FT has F1 score of 69 on N1(very large model with substantial training).

### Conv and attention based model(2020)
>[Notes](https://hackmd.io/GTpBvcSPQZOaYSG-FlC5VA)
- Uses Fpz -Cz EEG channel with window feature learning.
- Conv layers with attention mechanism and weighted loss is used.
- Sleep edf accuracy, MF1 scores are 93.7, 84.5.
- F1 score for N1 is 52.5.
- Sleep edfx accuracy, MF1 scores are 82.8, 77,8.
- F1 score for N1 is 47.1.

### SalientSleepNet(2021)
>[Notes](https://hackmd.io/_mh2X5fFQDCENtWo0Y3jGA)
- Uses raw EEG and EOG channels.
- Uses a U² structure with extraction and attention modules.
- Sleep edf-39 accuracy is 87.5 and F1 score for N1 is 56.2.

### TinySleepNet(2020)
>[Notes](https://hackmd.io/82LMwc_rRH2TOu3MscWCcg)
- Uses raw single channel EEG(F4-EOG/Fpz-Cz).
- Uses small conv layers with unidirectional LSTM.
- Sleep-edf(Fpz-Cz) accuracy, MF1 scores are 83.1, 78.1.
- F1 score of N1 is 51.

### XsleepNet(2021)
>[Notes](https://hackmd.io/QmZeJPQmRvOmI-mwQM6BMQ)
- Uses both raw signals and time frequency images.
- Has a CNN architecute for raw signal and bidirectional LSTMs for time frequency images.
- Has two variants based on certain assumptions.
- XsleepNet2, XsleepNet1 have 86.3, 86 accuracies respectively on sleep-edf dataset.

### Joint Classification and Prediction(2019)
>[Notes](https://hackmd.io/rjpXN3luTpmXBgIS5vkZGw)
- Uses time frequency images as input.
- Uses a CNN with multisoftmax layer and probability aggregation methods.
- Sleep edf and MASS accuracy using only EEG is 81.9 and 78.6% respectively.

### EEG Sleep Stages Classification Based on Time Domain Features and Structural Graph Similarity(2016)
>[Notes](https://hackmd.io/_a40zGJeQFuJFQKMo2Qyjg)
- Uses raw EEG signals and maps it into a graph.
- Features are extracted from the graph and fed into classifier based on K-means clustering algorithm.
- Sleep edf accuracy is 95.15%.
- Only comparison shown is with SVM method.

### Ensemble SVM Method for Automatic Sleep Stage Classification(2018)
>[Notes](https://hackmd.io/U1eT3b8wQA2VEd0fHc_XnQ)
- Uses MSPCA for preprocessing of EEG signal.
- Uses DWT for feature extraction and a modified SVM for classification.
- Sleep edf accuracy is 91.1%.

### GraphSleepNet(2020)
>[Notes](https://hackmd.io/T6QMFmiAT2O5oey5cJLC1w)
- Uses bandpass filters to preprocess EEG signals.
- A graph structure is used to learn the features and uses a spatial temporal graph convolution network for classification.
- MASS-SS3 used with accuracy 88.9 and N1 has F1 score of 60.3.

## Audio Classification

### CNN ARCHITECTURES FOR LARGE-SCALE AUDIO CLASSIFICATION(2017)
>[Notes](https://hackmd.io/yi1GXek6Sj2FpnTZNjnjxQ)
- Comparison of different Image classifers on audio classification.
- Uses log mel bin and fourier transform on raw audio.
- Large amount of training required on a large dataset for good accuracy.