Abstract— Epilepsy is one of the chronic brain disorders that affect the quality of life and well-being of millions of people around the globe. It is characterized by excessive electrical activity of the brain’s cells that usually leads to recurrent seizures. Accurate, efficient, and robust techniques suitable for recent Internet of Medical Things (IoMT) devices to detect, classify, and diagnose epileptic seizures in a challenging multi-classification scenario and noisy environment are of paramount importance. Electroencephalograph (EEG) signals recorded even from the surface of the brain suffer from contaminated artifacts and noise from various sources, such as from EOG and EMG for eye-blinks and muscle artifacts, respectively. This work aims to address the challenges of multi-class classification and automatic seizure detection in intracranial EEG signals by developing a detection system suitable for real-world clinical settings. To achieve this, this work uses an effective feature extraction technique and efficient seizure detection methods based on a recent big data resource, along with advancements in deep machine learning techniques, to propose and develop robust hybrid models that combine conventional machine learning techniques and deep learning architectures to increase the performance of epileptic detection systems to levels that are close to acceptable for real-world applications. Firstly, a robust computationally efficient technique that characterizes different types of seizures with high precision and low latency of its onset was proposed. The system relies on an effective and low in complexity feature extraction approach based on the proposed advanced time-frequency Fourier Basel series Expansion based Flexible Time-Frequency Analytic Wavelet Transform (FBSE-FTFAWT) that extracts notable features associated with EEG seizure signals in a time-effective manner. Secondly, two noise robustness seizure detection techniques were developed to address the research question: can the hidden patterns in artifact-induced epileptic EEG data be identified and characterized? Stacked Auto Encoder based Support Vector Machine (SAE-SVM) and Deep Belief Network based Support Vector Machine (DBN-SVM) as hybrid classifiers are proposed with a novel feature extraction to classify various seizure and non-seizure class combinations. The proposed optimized SVM classifier, FBSE-FTFAWT /SAE-SVM, shows better detection accuracy, sensitivity, specificity, precision, and F1-score of 99.7%, 99.6%, 99.6%, 99.7%, and 99.6%, respectively, over the other two proposed models and the state-of-the-art methods in the literature.