Signal processing and classification based on EMG data gathered from lower limb
The raw data used is published by Lencioni et al. and can be found from this link [1]
After preprocessing of the data followed by a feature extraction procedure, the main goal is to compare several algorithms by their performance.
Preprocessing includes filtering, noise elimination (exponential moving average) and interpolation/extrapolation of the raw data.
Feature extraction procedure includes many different time and frequency domain features used in literature. Some of them are mean absolute value (MAV), Mean Absolute Value Slope (MAVS), Simple Square Integral (SSI), Root Mean Square (RMS), Waveform Length (WL), Frequency Median (FMD), Modified Median Frequency (MMDF) and so on.
Multiclass classification and binary classification (( Step-up vs The Rest) and (Step-Down vs The Rest)) are aimed to be accomplished. Class imbalance problem persists, so SMOTE technique might be used to prevent the problems based on that.
Some of the planned models to use in this project are ANN, LSTM, DNN, SVM, Random Forest, Logistic Regression, LightGBM and XGBoost.
Scoring for binary classification will be based on accuracy metric and ROC curve. For multiclass classification log-loss metric will be used.
The progression is published continuously
References
[1] Lencioni, T., Carpinella, I., Rabuffetti, M. et al. Human kinematic, kinetic and EMG data during different walking and stair ascending and descending tasks. Sci Data 6, 309 (2019). https://doi.org/10.1038/s41597-019-0323-z