JoaoESmoreira

• MATLAB is used as the main development environment.
• Neural Network Toolbox functions such as feedforwardnet, layrecnet, and trainNetwork are employed.

The assignment focuses on the following learning goals:

1. Build, train, and test multilayer neural networks for classification of large datasets.
2. Build, train, and test dynamic neural networks (with delays) for multidimensional time series prediction.
3. Apply feature reduction techniques using autoencoders.
4. Configure, train, and test Convolutional Neural Networks (CNN) for multiclass classification.
5. Configure, train, and test Long Short-Term Memory (LSTM) networks for time series classification.

Epilepsy affects approximately 1% of the population, and around 30% of patients are resistant to treatment. EEG (Electroencephalogram) signals provide information about the brain’s electrical activity, which can be used to identify patterns associated with seizure states.

The goal is to classify EEG signals into three main brain states:

• Interictal: normal brain activity
• Preictal: the phase before a seizure
• Ictal: seizure occurring (including postictal phase)

The dataset contains EEG-derived features extracted from patients in the European Epilepsy Database (FP7 EPILEPSIAE). Each EEG segment is divided into 2-second windows with 50% overlap, resulting in one 29-dimensional feature vector per second.

Each group works with data from two patients:

• Odd-numbered groups: Patients 44202 and 63502
• Even-numbered groups: Patients 54802 and 112502

The project involves multiple machine learning models and techniques:

• Multilayer Perceptrons (MLP) for shallow classification.
• Dynamic Neural Networks (e.g., layrecnet) to include temporal memory.
• Autoencoders for dimensionality reduction of the 29-feature space.
• Convolutional Neural Networks (CNN) for image-based classification using 29x29 EEG feature maps.
• Long Short-Term Memory (LSTM) networks for sequence learning.

Data preprocessing includes:

• Normalization and class balancing (interictal, preictal, ictal)
• Temporal ordering of sequences
• Creation of target matrices with one-hot encoding for the three classes.

The models are evaluated using:

• Sensitivity (SE) — ratio of correctly predicted seizures.
• Specificity (SP) — ratio of correctly identified non-seizure states.

Both metrics must be maximized to ensure clinical relevance.

• MATLAB is used as the main development environment.
• Neural Network Toolbox functions such as feedforwardnet, layrecnet, and trainNetwork are employed.
• GPU and parallel computing are supported for faster training.
• A Graphical User Interface (GUI) is required to:
  • Select the neural network type (shallow/deep).
  • Choose datasets for training, validation, and testing.
  • Display training and testing results with sensitivity and specificity.

The final submission must include:

• Report file: ML2023EpiReportPLxGy.pdf
• Code and trained networks: ML2023EpilepPLxGy.zip

Each .m file must include a header describing its functionality and authorship.

The EEG features correspond to normalized spectral power in 29 frequency bands, ranging from 0.5 Hz to 512 Hz.

• European Epilepsy Database (FP7 EPILEPSIAE): https://www.epilepsiae.eu
• MATLAB Neural Network Toolbox (R2023b)
• ADC, DEI-FCTUC Machine Learning Course Materials