Find any frequency at any channel that predicts theta (evoked or induce) power at FCz or are predicted by theta at FCz: 1. Directly using power data 2. Not separating evoked/induced power which is averaged over trials, but using total power in trial by trial sense. Note: need to distinguish between conditions, especially between reward and no reward.
Try to find the relationship between EEG data recorded after feedback, and if there is anyway we can use it to predict behavior (choice and/or reaction time). Theta is involved somewhere, but perhaps we can find another frequency at another electrode that can predict whether the subject will make a left choice vs a right choice on the next trial, and whether these oscillations can predict whether the subject will win-stay (following positive feedback) or lose-shift (following negative feedback).
Private dataset collected via a 36-electrode EEG dataset from 135 college-age students. Each subject was sampled 200 times at a 250 Hz sample rate.
Initial data shaped as (135, 120, 1250) matrix. * 135 subjects * 120 frequencies * 1250 time samples
1250 time samples: 1. Button Press (1.5 seconds) - 371-372 time step 2. Feedback Onset (~2.5 seconds, 1 second after button press) - starts at 626 time step 3. Next trial (~3.5 seconds, 1 second after Feedback Onset) - starts at 888 time step Therefore, in order to use the feedback onset we use time steps 626-888.
Because the data was raw, a substantial amount of [preprocessing][preprocess.py] needed to be done. 1. Make the time data stationary 2. Apply a Box Cox transformation 3. Calculate difference to ensure stationality
Once this was done, the correlations between frequencies was compared to the theta frequency and the highest correlation was chosen. This was saved as testing/training datasets.
Various types of models were attempted. In addition to the currently working [models][train_ml.py] a transformer and LSTM-FCN were under development at the time of this writing. 1. RNN 2. GRU 3. LSTM
1. Clone repo - required libraries are stored in requirements.txt
2. Uncomment the model you wish to use in [train_ml][train_ml.py]. Make sure to uncomment the corresponding path as well.
3. Run the command `python train_ml.py`
4. Edit [eval_ml][eval_ml.py] to look for the correct model path
5. Run the command `python eval_ml.py`
Overall, the baselineGRU model performed the best in predicting targeted values where RNN performed the worst. The graphs are shown below:
The RNN Testing Dataset Predicition Versus Targets
The RNN Training Dataset Prediction Versus Targets
The GRU Testing Dataset Predicition Versus Targets
The GRU Training Dataset Prediction Versus Targets
The LSTM Testing Dataset Predicition Versus Targets
The LSTM Training Dataset Prediction Versus Targets
