Neurological disease detection from eeg signals with 1d and 2d cnn deep learning algorithms

In this thesis, the detection of Epilepsy, Alzheimer's, and Parkinson's diseases from 1D and 2D spectrogram-converted signals using deep learning was performed. Due to the complexity of EEG signals, an innovative approach is required for accurate neurological disease detection. Traditional methods often treat EEG data as a one-dimensional time series, overlooking significant frequency information in the signals. Our study proposes a methodology that converts 1D EEG signals into 2D spectrograms using the short-time Fourier transform (STFT) to address the limitations of one-dimensional data. This transformation captures frequency patterns, enhancing the analysis and providing a more comprehensive understanding of neural activity. In epilepsy seizure detection, the 1D CNN model accurately classified seizures and non-seizure instances with 98,9% accuracy, 98,9% sensitivity, 99,2% specificity, and an F1 score of 99,0%. When transformed into the 2D format, the CNN's accuracy slightly decreased to 98,5%, but sensitivity was 98,6%, specificity 98,4%, and the F1 score was 98,4%. For Alzheimer's disease detection, 1D EEG signals achieved 72,5% accuracy, 69,4% sensitivity, 75,6% specificity, and an F1 score of 71,8%, whereas 2D spectrograms achieved 82,6% accuracy, 76,9% sensitivity, 87,1% specificity, and an F1 score of 79,5%. In Parkinson's disease detection, 1D EEG signals achieved 74,3% accuracy, 69,2% sensitivity, 78,7% specificity, and an F1 score of 71,4%, while 2D spectrograms achieved 81,2% accuracy, 83,6% sensitivity, 79,2% specificity, and an F1 score of 79,9%. In conclusion, while 1D signals have shown greater success in epilepsy detection, 2D spectrograms provide higher accuracy, sensitivity, specificity, and F1 scores in detecting Alzheimer's and Parkinson's diseases. These findings demonstrate the increased effectiveness of using 2D spectrograms in the diagnosis of neurological diseases.