EEG-based schizophrenia diagnosis using deep learning with multi-scale and adaptive feature selection

Schizophrenia is a chronic and severe mental illness that significantly impacts the daily lives and work of those affected. Unfortunately, schizophrenia with negative symptoms often gets misdiagnosed, relying heavily on the clinician's experience. There is a pressing need to develop an objective and effective diagnostic method for this specific type of schizophrenia. This paper proposes a new deep-learning method called Cascaded Atrous Convolutional Network with Adaptive Weight Fusion (CA-AWFM) for classifying schizophrenia from electroencephalogram (EEG) data that combines cascaded networks with atrous convolutions and an adaptive weight fusion module (AWFM). This is because schizophrenia involves intricate and subtle brain wave patterns that make it difficult to detect the disorder from EEG signals. As such, our model uses an "atrous" convolution operation to extract multi-scale temporal information and a cascade network structure that progressively improves the attribute representations across layers. For classification purposes, AWFM enables our model to modify the importance of features dynamically. We evaluated our technique using a publicly available dataset of EEG recordings acquired from patients who have schizophrenia and everyday individuals. The proposed model has significantly outperformed existing methods with a 99.5% accuracy rate. With the help of atrous convolutions, local and global dependencies within the EEGs can be effectively modeled in this way. At the same time, AWFM makes flexible prioritization of characteristics possible for improved classification performance. With such impressive figures achieved, it can be concluded that our approach should be considered as accurate enough for routine clinical use in identifying schizophrenic patients early on so they can receive intervention measures on time or when diagnosed late, then dealt with appropriately.