This is biomedical data processing project by using MATLAB, you can have machine learning methods part done in Python or MATLAB, but signal processing should done in MATLAB at least.
The datasets contains 36 subjects EEG recordings (edf files) before and during the performance of mental arithmetic tasks that classified into good and bad performance. (The data and information you can see here, including published article: https://archive.physionet.org/physiobank/database/eegmat/).
Objectives:
- Analyze EEG features change from pre-task to during-task states in mental arithmetic performance by two signal processing techniques (FFT-based spectra, time-varying AR model).
- Explore how these changes differ between subjects with good versus bad performance by using two machine learning methods (ANN, GMM) to predict performance quality.
- Find accuracy of each combination of the signal processing approaches and machine learning models.
- Determine which combination has higher accuracy.
Detailed Instructions:
First, choose the most important channel or least noisy channel to analyze from 21 channels for each EEG based on any published research that indicate this channel is most related to arithmetic tasks. So you just need to analyze one channel for each EEG, then apply FFT-based Spectra and time-varying AR modeling signal analysis techniques to extract relevant features (including: frequency bands, AR coefficients, variance, residues, etc) from the datasets, you can use PCA to reduce features; Evaluate whether these changes in features resulting from counting are different between the bad and good counters; Then, employ two of the machine learning methods(ANN and GMM) to classify the datasets into good and bad counters (if GMM is not working, you can use CNN or other ML methods). For training and testing of ML methods, please using 5-k fold cross validation, just make sure there are some kind of balance of good and bad counters data in each fold (since there are 24 good and 12 bad data). Then find the most accurate combination of signal processing technique and machine learning methods. Also, if possible, determine which features contribute the most to the classification.
You can have your report like this:
Due on April 15, 6 PM PST: Some preliminary results are due.
Brief overview of Project: Disclose any changes in Specific Aims, following preliminary efforts.
• Reminder of datasets used – example figures of sample data (time-series). Raw vs “processed” data.
• Signal processing methods used – preliminary results, parameters/features extracted.
• Machine learning methods used – preliminary results.
• Problems encountered – anticipated and unexpected? How will you address these problems?
• What do you plan to do next – what should we expect in the final presentation?
Due on April 20, 6 PM PST
All project should finished.