Reconstructed phase space portraits for detecting brain diseases using deep learning.

• Identify normal and diseased EEG signals using Reconstructed Phase Space (RPS) images. • Using RPS portraits to model the nonlinear dynamics of the EEG signal implicitly. • Detecting multiple brain diseases using RPS images with the comparison of CNN models. • Highest accuracies of 99.37% for schizophrenia, 96.63% for sleep disorders is reported. • RPS based ensemble model for multiple brain disease identification with 98.44% accuracy. Background and objective Abnormalities in the brain affect its functions, and is observed among people of all background. Automatic identification of these abnormalities will assist treatment. These abnormalities are observable in the electroencephalogram (EEG) signal which measures the electrical activity of the brain. The EEG signal is nonlinear and chaotic in nature. Automatic detection of certain diseases affecting the brain has been attempted using the spatial and temporal characteristics of the EEG employing signal processing, image processing and machine learning techniques. We propose the automatic identification of the abnormalities in the brain based on the depiction of its nonlinear chaotic nature. Methods Accurate detection across multiple brain diseases is a far more challenging problem given the dynamic nature of the EEG signal. We employ Reconstructed Phase Space (RPS) portraits to model these nonlinear dynamics. The EEG signal analysis of short segments results in RPS portraits that implicitly capture the dynamics. In this study, we examine the performance of CNN-based deep learning models as feature extractors from RPS portraits of EEG signals in achieving an improved screening of brain diseases, by classifying them as normal or diseased. Results The use of CNN architectures like Inception-v4 results in high accuracy (97.30%) in the identification of brain diseases (Epilepsy, Schizophrenia, and Sleep Disorder). A further improvement (98.44 ± 1.5%) in the disease identification is achieved using a voting classifier. Conclusions The non-linear dynamics and CNNs for learning salient features suitable for accurately identifying the brain disease in an end-to-end system is a novel contribution of this study. The high performance of the system makes it feasible to study its utility as a diagnostic support tool. [ABSTRACT FROM AUTHOR]