AJILE Data Readme

Data Organization

ECoG data

 

This is the voltage potential data from all electrodes over time stored in EDF format. A recommended viewer of EDF data is edfbrowser, and EDF formats can be loaded with PyEDFlib

or mne in python.

 

The ECoG data is split by day for each subject. The day is in reference to the date of surgical implantation of the electrodes, where day 1 is the day of the surgery. Due to the often poor patient conditions in the first couple of days after surgery, the data from these days are generally not used. Each file contains data from midnight to midnight for each day and if there is missing* or purged** data, the voltage is set to the physical minimum (-15654.1 uV) for that period. 

Joint Locations

Joint locations are stored in CSV files for every frame (30 frames per second) whenever available. The CSV file contains the following columns: Time (real time of frame***), Missing (Whether this video portion is missing), Purged (Whether this video portion has been purged), 7 joints. For each joint, the tuple contains (x-location, y-location, confidence). Generally, confidences above 0.25 can be trusted. When the data is missing* or purged**, the locations and confidence values are shown as -1 for that frame. 

Electrode Locations

Electrodes are rendered on a brain reconstruction in several views. In addition, the numerical 3D coordinates of the electrodes are also provided in a .mat file for each subject. 

Training, Development and Test data times used for movement initiation prediction

The actual training, development and test data used in the 2018 AAAI movement initiation paper is provided in a CSV file for each subject. The columns are as follows: Type (train, dev or test), mvmt( right arm movement->r_arm_1, no movement->mv_0, left arm movement->l_arm_1), day, time (both corresponding to data provided above). Please note that for each patient, we compared contalateral movement intiation so a single subject will not have both right arm and left arm movement in the data used. 

Helpful Code

https://github.com/BruntonUWBio/AJILE-tools is a toolbox in development (contibutions welcomed!) to easily access and visualize the AJILE dataset. It contains a Jupyter notebook that is helpful to begin loading and exploring the data. It also contains a skeleton animator script so that one can visualize the upper body joints in real time. 

Misc. Notes

*Missing->Sometimes, large (hours) or small(seconds) chunks are missing either from the EDF file, video file (thus joints file) or both due to recording issues at the clinical site or became missing later on. These are not common but users need to make sure that they do not accidentally take into account these times. 

**Purged->When the patient is engaged in some activity that may be considered embarrassing, we manually flagged these times and purged them from the joints and ECoG data. These times may have been used in the movement initiation paper however so ensure that when replicating results, purged data is not included in the analysis. 

***Real time of frame-> For each time stamp, we include up to microseconds but realistically, the time is accurate to within 100 milliseconds. As well, please disregard the date. The actual dates of recording have been stripped to further anonymize the patient data. 

Citation: Wang, NXR., Farhadi, A., Rao, R., & Brunton, B. (2018). AJILE Movement Prediction: Multimodal Deep Learning for Natural Human Neural Recordings and Video. In Proceedings of the 32nd AAAI Conference on Artificial Intelligence (AAAI).