Page 120 - AN-4-2

Advanced Neurology ML for EEG signal recognition

each representing an EEG examination where each row disabling a fraction of neurons during training. The final
represents a different patient. It includes data from various output layer utilized a sigmoid activation function to
channels across the scalp, capturing different frequency produce a binary classification output – epileptic or non-
bands. The dataset comprises 668 columns, of which epileptic. The DNN was trained using the Adam optimizer,
253 columns represent specific frequencies recorded at which dynamically adjusts learning rates to accelerate
different scalp locations. Additional columns provide convergence. The binary cross-entropy loss function was
statistical metrics (e.g., mean and standard deviation) for chosen to optimize classification accuracy. Training was
specific frequency bands at various times. Local sleep-wake conducted over 50 epochs with a batch size of 32, and early
transition band values are also recorded. The target column stopping was applied based on validation performance to
indicates whether the individual examined was epileptic prevent overfitting.
(1) or non-epileptic (0). This comprehensive dataset was
used to identify EEG features indicative of epilepsy and to 2.3.2. Convolution neural network
develop predictive models capable of screening patients for The CNN was designed to exploit the spatial and temporal
epilepsy. relationships within the EEG data. The model architecture
began with two convolutional layers, each equipped with
2.2. Preprocessing ReLU activation functions and configured with 32 and 64
The dataset was preprocessed to ensure suitability for filters, respectively. These layers were followed by max-
machine learning modeling. Initially, the data were loaded pooling operations to downsample feature maps and
into the Python environment and the target column was reduce computational complexity. Dropout layers, with
separated from the predictor variables. The dataset was a rate of 0.3, were included to enhance generalization
then divided into three subsets for training, validation, by preventing overfitting. The CNN further included a
and testing, with proportions of 70%, 20%, and 10%, flattening layer to convert the multi-dimensional feature
respectively. Stratified sampling was employed to preserve maps into a one-dimensional feature vector. This vector
the original class distributions in each subset, ensuring was processed by a series of dense layers, culminating in
that both epileptic and non-epileptic cases were adequately an output layer identical to that of the DNN. Training
represented. Predictor variables were standardized using for the CNN followed a similar protocol, employing the
a StandardScaler (class from the Python library scikit- Adam optimizer and binary cross-entropy loss function,
learn) to normalize the feature values. This step was with model evaluation occurring at each epoch to monitor
critical for neural network models to achieve consistent progress.
convergence during training. For the convolutional
neural network (CNN), the standardized predictor data 2.3.3. PyCaret machine learning
was reshaped to include a channel dimension, enabling The third approach involved the use of PyCaret, an
the model to interpret each feature as part of a temporal automated machine learning framework. PyCaret’s
sequence. classification module was configured to preprocess the
data, evaluate multiple machine learning algorithms, and
2.3. Machine learning models identify the best-performing model based on a variety
Three distinct machine-learning approaches were of metrics. Following the automated model comparison,
implemented and evaluated in this study. Each model the selected model underwent hyperparameter tuning
architecture was carefully designed to leverage the to optimize performance further. The final tuned model
unique characteristics of the EEG dataset and optimize was then evaluated on the test set, with key metrics such
performance for epilepsy prediction. as accuracy, precision, recall, and F1-score calculated to
assess its predictive capability.
2.3.1. Dense neural network (DNN)
The dense neural network (DNN) was constructed using 2.3.4. Model implementation and training details
the Keras framework. This model comprised an input All models were implemented in Python 3.8.10 using
layer configured to match the dimensions of the predictor TensorFlow 2.7.0 (Keras API) for neural networks
variables, followed by three hidden layers. Each hidden and PyCaret 2.3.10 for LightGBM. Hyperparameter
layer employed rectified linear unit (ReLU) activation optimization used: (1) manual grid searches to constrain
functions to introduce non-linearity and enhance the ranges, then (2) Keras Tuner (20 trials) for CNNs/DNNs,
model’s ability to capture complex patterns in the data. targeting kernel sizes (3 – 15 samples), dropout (0.2 –
Dropout layers with a rate of 0.3 were incorporated after 0.5), and L2 regularization (1e-4 – 1e-2). ReLU activation
each hidden layer to mitigate overfitting by randomly (hidden layers) and sigmoid (output) ensured non-linear

Volume 4 Issue 2 (2025) 114 doi: 10.36922/an.7941