This document discusses using a single channel EEG device to recognize emotions from EEG data. It collected data from 10 individuals labeled as stressed or relaxed. It preprocessed the raw EEG data using filters to isolate brain signals. It then used deep learning models including an LSTM network with and without attention to classify emotions. The LSTM with attention achieved 85% accuracy, which was an improvement over the LSTM without attention. Potential applications discussed include using EEG for stress reduction by customizing music, and emotion or word prediction. The document also discusses opportunities for future enhancements such as using convolutional layers or multi-modal networks incorporating additional physiological sensors.
5. CONSUMER EEG DEVICES
▪ 1 to 6 channels based on the device
▪ Dry electrodes makes it highly portable
▪ Wireless data collection
▪ Easy integration for programming
▪ Mostly focuses on Pre-Frontal cortex which is responsible for
▪ Attention
▪ Decision making
▪ Emotion management
▪ Planning
▪ Coordinating complex behavior
▪ Affordable
6. Presentation Disclaimer
▪ This presentation and the related paper are the works of individuals based on
personal interest. None of the organizations that these individuals work for are
responsible for the accuracy and completeness of the content or results. This work
and presentations are for information purposes only.
7. DATA COLLECTION
Collected Data:
▪ Single channel (FP1) EEG device
▪ Collected data from 10 individuals, multiple episodes
▪ Binary labels – ( Stressed, Relaxed )
▪ Only Raw signal is used for this experiment
Public Data:
▪ SEED
▪ DREAMER
▪ https://www.kaggle.com/phhasian0710/eeg-fpz-cz
Public data is used to pretrain the network and experimented in embedding the signals.
8. PRE-PROCESSING RAW EEG
▪ Raw EEG consists of signals from different sources
▪ Muscle movements
▪ Eye movements
▪ Blinks
▪ Apply FFT / Butterworth band pass filters to separate out brain signals.
▪ EEG Lab
▪ MNE Tools
▪ Scipy
▪ Baseline the signal
▪ EEG Patterns vary with Age and hence it needs to be considered as a feature.
9. DEEP LEARNING FOR EEG
▪ Pretrain the network with a public dataset (
we used one from Kaggle )
▪ Data from FP1 and FP2 ( public data) are
averaged to form FPZ
▪ Select pre frontal points and average it, if the
data contains multi channel
▪ Classifier trained to identify emotions like
“dominance”, “high arousal”, “like” etc.
▪ Data collected from Neurosky is at 512 HZ
▪ Each reading is of size 7500
▪ Raw signals, without FFT is provided as an
input
▪ Limited architecture to compensate low
volume
10. RESULTS
▪ Model Learnt to ignore eye blinks and other noises
▪ However pure LSTM model failed to focus on the right patterns
▪ Adding attention layer improved accuracy drastically.
Metrics LSTM without
attention
LSTM with
attention
Accuracy 55 % 85 %
F1 score 55 % 85 %
Recall 55 % 85 %
Precision 55 % 85.42 %
Average accuracy Relaxed Stressed
LSTM without Attention 43.75 62.5
LSTM with Attention 89.47 80.952
11. FUTURE ENHANCEMENTS
▪ Convolution layers to autofocus and reduce computational cost
▪ Signal Embedding is not considered however we believe it will improve performance
▪ More data collection with more subjects and standardized stimuli
▪ Multi Modal networks with inputs from ECG, GSR and EEG
▪ Time synchronized data consolidation from public sources.
12. POTENTIAL APPLICATIONS
▪ Custom music composer to reduce stress ( based on
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5840261/ )
▪ Emotion and word prediction from EEG