/myo-readings-dataset

Myo armband electromyographic readings dataset for various wrist gestures.

Primary LanguageJupyter NotebookMIT LicenseMIT

myo-readings-dataset

Myo armband electromyographic readings dataset for hibernation, flexion, extension, radial deviation, ulnar deviation, pronation, supination and fist gestures of the wrist.

Project structure

The readings from the right hand are located in the _readings_right_hand folder, while the readings from the left hand are located in the _readings_left_hand folder which both contain one folder per recording session. All session folders containig data from the same participant should be named by a random unique id (five digits), followed by a dash and session number, starting at 1 (for example 12345-1, 12345-2, etc.). Each session folder contains multiple files, one per each wrist gesture. These are named <label>.txt (e. g. 2.txt for extension, see Gesture labels below). Each recording session folder should contain at least eight files (for gestures 0-7). Three sessions are expected to be recorded on either the right or the left hand. The file itself is composed of multiple lines:

...
11,32,-3,-43,4,5,42,7,0
13,24,-5,12,43,42,12,1,0
123,121,-100,-88,-32,32,123,13,2
...

Each line represents the samples from the eight EMG channels on the Myo armband ([-128, 127], signed byte) as well as the label for the wrist gesture (class) in a given time moment, separated by commas. There is no comma at the end of the line and there should be no spaces anywhere in the file. The sampling frequency is approximately 200 Hz, as per Myo specifications. In this example, the first two lines represent hibernation (e. g. 0 at the end of the line), while the third line represents extension (e. g. 2). The EMG values are arbitrary in the example. There should not be an empty line at the end of the file.

Recording protocol

The Myo armband is placed on the thickest part of the (preferably) right forearm, with the LED pointing towards the dorsal (back) part of the hand. Each file contains one minute of recording and should be around 12 000 lines long (~200 Hz × 60 s = 12 000). The labels at the ends of the lines alternate between hibernation (0) and the gesture denoted in the file name every five seconds. The exception is the hibernation gesture, where the file should contain only hibernation labels (0) at the ends of the lines. Hibernation recordings contain various slight movements and motions of the hand that we would prefer to ignore during classification.

Gesture labels

  • 0: hibernation

  • 1: flexion

  • 2: extension

  • 3: radial deviation

  • 4: ulnar deviation

  • 5: pronation

  • 6: supination

  • 7: fist

Curated sessions

The curated.txt file contains a list of sessions that yield a decent intraparticipant accuracy accross most gestures when applying machine learning strategies to evaluate the dataset. Good intraparticipant accuracy indicates proper electrode placement and device orientation during individual recording sessions of a single participant. The curated sessions hence also yield a good interparticipant accuracy. If you intend to use the dataset for applying machine learninig strategies, it is recommended to use only the curated sessions.

TODO

  • Automatic curated.txt generator (curated.txt is outdated)