Ryan Gabriel

Epileptic seizure onset detection from electroencephalography (EEG) signals is essential for enhancing patient safety and enabling timely clinical interventions. While prior research has explored various feature extraction methods and deep learning architectures, few studies have systematically compared the effectiveness of different time-frequency representations (TFRs), particularly for onset-centered EEG segments. In this study, we systematically evaluate the impact of three TFRs on seizure onset detection using deep learning models. We analysed EEG data from the TUH seizure corpus dataset using 20-second segments that include 10 seconds of background followed by 10 seconds of seizure onset. Initially, EEG signals were pre-processed and transformed into spectrograms using the three TFR methods: short-time Fourier transform (STFT), continuous wavelet transform, and Mel spectrogram. Further, the spectrograms were processed via two channel encoding strategies: RGB spectrogram encoding and grayscale spectrogram encoding. Five attention-based deep learning models, including long short-term memory with attention (ATT-LSTM), transformer, convolutional neural network-transformer, vision transformer, and shifted window transformer were trained and evaluated under patient-wise 5-fold cross-validation. We conducted both binary (seizure vs. non-seizure) and multi-class (focal, generalized, background) classification tasks. Our results show that STFT consistently outperformed the other TFRs across all models, with the ATT-LSTM model achieving the highest mean accuracy of 82.37% in the binary task. RGB spectrogram encoding also yielded superior performance compared to the independent approach. However, the transition to multi-class classification proved more challenging, with a notable drop in accuracy. These findings demonstrate that early seizure detection can be achieved with high reliability using STFT-based spectrograms, especially when combined with composite spatial encoding. However, distinguishing seizure types at onset remains a complex task, suggesting the need for richer contextual modeling in future work.

Social media was introduced as a way to communicate and interact with people. The introduction of social media bots tampered with the sovereignty of social media as a whole. Bots alter the social media experience in many ways, including swarming comments and manipulating the algorithm to favour the creator of the bots. This is a significant issue for social media platforms and must be addressed immediately. While earlier methods experimented with Ensemble classifiers such as XGBoost, Random Forests, and basic NLP, modern approaches include the use of Graph Neural Networks (GNNs) along with Heterogeneous Graph Networks, Graph Attention Networks (GAT), and Graph Convolutional Networks (GCN) for optimal performance, generalising across multiple datasets. In This Paper, we discuss the taxonomy of bots, the ways they affect social media and influence public opinion, and the various approaches for Social Media bot detection on datasets such as TwiBot-20, TwiBot-22, Cresci-2017, MGTAB, and the Graph-Based models used. We also examine evaluation metrics such as Recall, Precision, F1 score, current limitations, and future scope.