DEEP ARCHITECTURES FOR SPATIO-TEMPORAL SEQUENCE RECOGNITION WITH APPLICATIONS IN AUTOMATIC SEIZURE DETECTION

Abstract

Scalp electroencephalograms (EEGs) are used in a broad range of health care institutions to monitor and record electrical activity in the brain. EEGs are essential in diagnosis of clinical conditions such as epilepsy, seizure, coma, encephalopathy, and brain death. Manual scanning and interpretation of EEGs is time-consuming since these recordings may last hours or days. It is also an expensive process as it requires highly trained experts. Therefore, high performance automated analysis of EEGs can reduce time to diagnosis and enhance real-time applications by identifying sections of the signal that need further review.Automatic analysis of clinical EEGs is a very difficult machine learning problem due to the low fidelity of a scalp EEG signal. Commercially available automated seizure detection systems suffer from unacceptably high false alarm rates. Many signal processing methods have been developed over the years including time-frequency processing, wavelet analysis and autoregressive spectral analysis. Though there has been significant progress in machine learning technology in recent years, use of automated technology in clinical settings is limited, mainly due to unacceptably high false alarm rates. Further, state of the art machine learning algorithms that employ high dimensional models have not previously been utilized in EEG analysis because there has been a lack of large databases that accurately characterize clinical operating conditions. Deep learning approaches can be viewed as a broad family of neural network algorithms that use many layers of nonlinear processing units to learn a mapping between inputs and outputs. Deep learning-based systems have generated significant improvements in performance for sequence recognitions tasks for temporal signals such as speech and for image analysis applications that can exploit spatial correlations, and for which large amounts of training data exists. The primary goal of our proposed research is to develop deep learning-based architectures that capture spatial and temporal correlations in an EEG signal. We apply these architectures to the problem of automated seizure detection for adult EEGs. The main contribution of this work is the development of a high-performance automated EEG analysis system based on principles of machine learning and big data that approaches levels of performance required for clinical acceptance of the technology. In this work, we explore a combination of deep learning-based architectures. First, we present a hybrid architecture that integrates hidden Markov models (HMMs) for sequential decoding of EEG events with a deep learning-based postprocessing that incorporates temporal and spatial context. This system automatically processes EEG records and classifies three patterns of clinical interest in brain activity that might be useful in diagnosing brain disorders: spike and/or sharp waves, generalized periodic epileptiform discharges and periodic lateralized epileptiform discharges. It also classifies three patterns used to model the background EEG activity: eye movement, artifacts, and background. Our approach delivers a sensitivity above 90% while maintaining a specificity above 95%. Next, we replace the HMM component of the system with a deep learning architecture that exploits spatial and temporal context. We study how effectively these architectures can model context. We introduce several architectures including a novel hybrid system that integrates convolutional neural networks with recurrent neural networks to model both spatial relationships (e.g., cross-channel dependencies) and temporal dynamics (e.g., spikes). We also propose a topology-preserving architecture for spatio-temporal sequence recognition that uses raw data directly rather than low-level features. We show this model learns representations directly from raw EEGs data and does not need to use predefined features. In this study, we use the Temple University EEG (TUEG) Corpus, supplemented with data from Duke University and Emory University, to evaluate the performance of these hybrid deep structures. We demonstrate that performance of a system trained only on Temple University Seizure Corpus (TUSZ) data transfers to a blind evaluation set consisting of the Duke University Seizure Corpus (DUSZ) and the Emory University Seizure Corpus (EUSZ). This type of generalization is very important since complex high-dimensional deep learning systems tend to overtrain. We also investigate the robustness of this system to mismatched conditions (e.g., train on TUSZ, evaluate on EUSZ). We train a model on one of three available datasets and evaluate the trained model on the other two datasets. These datasets are recorded from different hospitals, using a variety of devices and electrodes, under different circumstances and annotated by different neurologists and experts. Therefore, these experiments help us to evaluate the impact of the dataset on our training process and validate our manual annotation process. Further, we introduce methods to improve generalization and robustness. We analyze performance to gain additional insight into what aspects of the signal are being modeled adequately and where the models fail. The best results for automatic seizure detection achieved in this study are 45.59% with 12.24 FA per 24 hours on TUSZ, 45.91% with 11.86 FAs on DUSZ, and 62.56% with 11.26 FAs on EUSZ. We demonstrate that the performance of the deep recurrent convolutional structure presented in this study is statistically comparable to the human performance on the same dataset.