This project is an effort to use machine learning to understand the mental workload generated on an individual while performing a task and classifying the level of cognition induced in an individual.
The goal is to use various data processing techniques and ML architectures to preserve both spacial and time information in the classification of EEG data.
For a more concise and visually pleasing presentation of this project, please see the included PDF. (BTP report.pdf)
*THE EEG DATA HAS BEEN CAPTURED @BCI-HCI LAB, COMPUTER SCIENCE & ENGINEERING DEPARTMENT, IIT KHARAGPUR, INDIA.
EMOTIVE - Epoc+
SENSOR COUNT - 14+2 references
SENSOR - AF3, AF4, F3, F4, FC5, FC6, F7, F8, T7, T8, P7, P8, O1, O2
SAMPLE RATE - 128 HzCritical task and cognition-based environments, such as in military and defense operations, aviation user-technology interaction evaluation on UI, understanding intuitiveness of a hardware model or software toolkit, etc. require an assessment of how much a particular task is generating mental workload on a user. This is necessary for understanding how those tasks, operations, and activities can be improvised and made better suited for the users so that they reduce the mental workload on the individual and the operators can use them with ease and less difficulty.
We propose an approach to automatically evaluate the task complexity perceived by an individual by using electroencephalogram (EEG) data of a user during operation.
The data are in the form of csv files with raw waveform signals from 14 probes places around the scalp. The sampling rate is 128 hz, which allows for frequency analysis up to ~60 hz. Each of 8 subjects participated in model building using SolidWorks
Each subject underwent 3 trials. Each trial included the following steps-
- The first 5 seconds of procedure was quite; the subject was Idle.
- At the beginning of 6th second, the subject started building the model.
- Finally, after a given time limit and 5s, idle time after the experiment indicates the end of the trial.
The link to Level 1, Level 2, Level 3 data : CLICK HERE
Most of the EEG signals recorded are contaminated by artifacts and do not represent actual brain signals. Artifacts can be significantly larger than the EEG signal, and from various sources like muscle contraction or electromyogram (EMG), eye movement or electrooculogram (EOG), heart activity or electrocardiogram (ECG) and power line noise.
. In this work, the recently developed technique called FORCe (Fully online and automated artifact removal) for brain-computer interfacing method has been used. Following the removal of artifacts, the data is further processed.
From the description of the trials, the readings corresponding to the respective mental load of a particular task from all the subjects were divided accordingly. Three trials were taken for each individual, for each level of mental workload and were clubbed together. Each of these readings is further divided into 1-second epochs as shown in the figure below (see Figure ). The sampling frequency of the device used is 128 Hz, so for an epoch of 1 second, we got 128 samples.
The filtered signals obtained for each epoch are of high dimension. It is important to reduce the complexity of such high dimension signals. We extracted the features from the processed EEG data, which gives information about distinct components of the EEG data. It is an important step, as its extraction is needed for understanding various features of the EEG data by the machine learning algorithm. We extracted a total of 52 features which gives a lot of information about the type of brain waves we are dealing with and helps machine learning algorithms to better understand the data.
The features that are extracted are then normalized to bring them to a common range. This optimization helps in reduction of inter-subject variability. Here, the extracted features are mean-normalized
Overfitting and dimensionality curse can be minimized using feature selection and optimization. Features which are strongly correlated to the target variables, which in our case is the task type are selected for classification. Feature selection becomes important because it not only decreases the number of features for further processing but also increases the computation speed since the machine learning algorithm has to deal with feature space of low dimension. Feature Selection Method Used:
We used the three different types of feature selection techniques i.e. Trees Based Feature Selection (Extra Trees Classifier), XGBoost Feature Selection and Correlation-based Feature Selection. We were able to get the most important features for each specific feature selection and feature important techniques. Based on the results, we selected top 10 features for each selection technique. The results for each of them are as follows:
Top 10 most important features from different feature selection technique.
| Extra Trees | XGBoost | Correlation |
|---|---|---|
| Wavelet Detailed | STD AR | Wavelet App. Entropy |
| Wavelet Detailed Energy | Wavelet Detailed STD | Hjorth_activity |
| Wavelet App. Entropy | Variance of V to V slope | Variance of V to V slope |
| Auto Regressor | Wavelet Appx. STD | Wavelet Detailed Energy |
| Wavelet Appx. STD | Kurtosis | Wavelet Appx. STD |
| Variance of Vertex to Ver slope | Hjorth_mobility | FFT Beta Max Power |
| Delta/ Theta | Wavelet App. Entropy | 1st Difference Max |
| Wavelet Appx. mean | Delta/ Theta | FFT Alpha Max Power |
| FFT Delta Max Power | Wavelet App. Energy | Coefficient of Variation |
| Delta/ Alpha | Wavelet Appx. Mean | FFT Theta Max Power |
Comparisons were done in the classification of the target variable using all features and only the selected features. We found that although the selected features were comparatively less accurate but they proved to classify the target value to a great extent. Below is the table containing the overall accuracy of different machine learning models when all features were used and when only a limited number of selected features were used.tc
One of the initial goals of the project was to explore the feasibility of wireless data acquisition devices in task evaluation and MWL assessment. It is evident that that approach used in the work can be used to understand the complexity of a task and this information can be used by companies and industries to make their software, hardware and UI toolkits better and more intuitive for the user. From the results obtained it can be observed that XGBoost classifier gives the best accuracy in comparison to other machine learning algorithms. We hope that this study would be helpful in future to explore and devise new methods for studying and understanding the task, their complexity and the mental workload required in its operation, helping in improving the usage and user interface of various software and hardware toolkits.
Level-1 Data - Contains csv for the EEG readings of 8 subjects for task 1
Level-2 Data - Contains csv for the EEG readings of 8 subjects for task 2
Level-3 Data - Contains csv for the EEG readings of 8 subjects for task 3
BTP_Code.ipynb - Contains the project code
Normalsiedfeatures.csv - contains the data after applying Force Algorithm
features.csv - Contains the fully feature engineered dataset of the EEG signalsVishal Anand- Department of Mechanical Engineering, IIT Kharagpur
Zaki Ahmed - Department of Mechanical Engineering, IIT Kharagpur
Sr Sreeja - Department of Computer Science and Engineering, IIT Kharagpur