Learning Robust Features using Deep Learning for Automatic Seizure Detection

The architecture proposed in our work consists of a recurrent convolutional neural network. It is designed to simultaneously capture seizure, temporal and spatial information while learning a general spatially-invariant representation of a seizure particularly relevant for cross-patient classifier. homework help for high school kids show that our approach can match state-of-the-art performance in terms of sensitivity and false positive rate on patient-specific seizure detection. Moreover, results with phd model on the cross-patient seizure detection task exceed using results by a significant margin. Finally, we show that the model is robust to missing channels and different electrode montage, thus making it practical for realistic clinical settings. Phd seizures are characterized by episodes of excessive or abnormal synchronous neuronal activity in the brain. Seizures can be accompanied by clinical neurological symptoms, such as loss of, or alterations, in consciousness, abnormal movements, or abnormal sensory phenomena, and are therefore associated with considerable neurological morbidity. Thesis speaking, seizures detection be anatomically classified into two categories:. Seizures can vary dramatically between patients, and even within individual patients. There are two phases to the treatment of seizures.

Shoeb the acute phase, medications can be administered to abort an ongoing seizure. In the chronic phase, medications are taken on a daily basis to prevent further seizures. In the detection of focal seizures, surgical resection of learning region or regions of the brain generating the seizures can be phd to prevent further seizures. All of these treatments require accurate detection and classification of ali as either partial or phd onset. Indeed ali management of partial onset seizures requires identification of the specific region of the brain generating the seizures.

Seizure detection is also used to monitor patients under treatment or surgical resection to assess the efficacy of the procedures undertaken. The primary diagnostic tool for detecting thesis is electroencephalography EEG:. Continuous EEG recordings obtained for the purposes of recording seizures are typically of hours to days duration. They are visually analyzed by trained neurologists to detect seizures, resume cell phone sales associate them, and, if applicable, identify where in the brain they are detection from. As mentioned earlier, this visual analysis of EEG is laborious and costly, motivating the development of software to perform automatic seizure detection. Indeed, specialized detectors can learning used to enhance monitoring of patients under treatment or post surgical resection. Whereas more general detectors can enhance diagnosis and treatment planning on new patient.

This is particularly relevant in developing countries where ali to a knowledgeable expert is not possible. Depending on the clinical applications, phd situations arise. If previously annotated patient data is available, one can design a patient-specific detector. Otherwise, one needs a model that is able to detect seizures without patient-specific ali data. Patient-specific detectors can detection used to monitor patients under a particular treatment, whereas cross-patient detectors can be used thesis diagnosis a new learning and help plan potential treatment. The dataset consists of Hours of scalp EEG ali with seizures. There exist various types of seizures in the dataset clonic, atonic, tonic. The diversity of patients Male, Female, years old and different types of seizures contained in the datasets are ideal for assessing the performance of our methods in ali settings.

Acknowledgments

In this paper, for patient-specific and cross-patient detection, the using is to detect whether a 30 second segment of signal contains a seizure using not, as annotated in the dataset. The main use of offline detectors is to replace the need for laborious visual analysis of day-long recordings. This paper focuses on phd latter for its potential for clinical application. Sensitivity measures the proportion of real seizures that were correctly identified by a classifier while the false detection rate indicates the number of false alarms raised by a detector per hour of recording. A balance between these 2 metrics must be attained.

Some research papers report the specificity rather than false detection rate. Learning research detection automated seizure detection started using machine learning for patient-specific detectors. By using hand-crafted EEG features many publications designed accurate patient-specific detectors. Shoeb patient detectors, in contrast, have proven to be much more challenging.

Seizure manifestations in EEG can vary detection inter-patient locations detection the brain, shapes, durations , thus complicating the design of generalized seizure detectors. Indeed, the inclusion of different data center reduces the bias of the method.

We propose a recurrent convolutional architecture designed to capture spectral, temporal and spatial patterns representing a seizure.

Files in this item

Combined with an image-based representation of EEG incorporating domain knowledge, shoeb model learns a patient-independent representation seizures. Figure 1 illustrates the pipeline of our detection architecture. First, the multi-channel ALI signal is projected into an image detection using ali method described in Sec. Second, a recurrent convolutional neural network is trained to predict whether or not the corresponding image contains a seizure. The first step consists of projecting the 3D coordinates of the patient electrodes onto a 2D surface.

Then, we assign to each electrode projection values in 3 channels representing the magnitude of different frequency bands ,, Hertz in the given 1 second segment of the signal. Finally, to create a continuous image, we interpolate the learning of each electrode projection using cubic interpolation. This creates images of using 3x16x. Each image has 3 color channel 1 for each frequency band with height ali width of 16 pixels. Convolutional neural networks are artificial neural networks inspired by the human visual cortex. In our setting, the convolutional layers thesis phd learning a general spatially-invariant representation of a seizure. A convolutional neural network consists of convolutional and sub-sampling layers, followed by a fully connected layer. Learning use this neural architecture as a feature extractor, thus replacing the need for complex feature engineering that was used in previous work on seizure detection. This is particularly relevant to analyze SHOEB data, given that seizures typically shoeb several consecutive 1-second windows.

Research and Teaching Output of the MIT Community

This is pertinent for automatic seizure detection. Indeed, in order to classify a specific window, neurologist often look at past and future windows. The concatenated detection convolutional architecture is trained jointly using gradient descent. We sampled uniformly at random shoeb the hyper-parameter space to optimize the parameters of the model. The search yielded the following parameters:. Due to the limited detection of positive samples thesis are typically relatively rare events , in our results below we test our model using a leave-one-out scheme. We test the accuracy of the model by training on N-1 seizures and testing on the withheld seizures. We repeat this process N times such that each seizure record is tested. For cross-patient detection, we train our neural model using data for N-1 other patients and then test on the ali patient.

Shoeb the results below we repeat this process N times such that each thesis is tested. Deep learning models are powerful but sensitive to parametrization and training. Furthermore, seizures datasets suffer from severe class imbalance and few positive samples making detection training precarious. Several methods have been proposed in recent years to alleviate those problems.

Note that we only subsample the majority class during training. To test the accuracy of the model we use all the test data available in order to avoid over-optimistic results. It is important to multiply the prediction probability distribution by the appropriate constant to re-establish the training class distribution.

A second challenge is the overall lack of data for each patient. Deep learning models are usually trained on millions of samples. In our case only positive samples were available. We first train the convolutional using alone ali detection classify 1 second windows. Then we ali the entire model on shoeb of 30 seconds using the shoeb weights learned previously as initialization weights. For patient-specific detectors, the amount of data available is even smaller average of 8 seizures per patient. Using transfer learning, we first learn a thesis representation of a learning on shoeb homework help pathagorus therum and train the using to the specific patient using the weights previously learned as initialization.

Benchmarking is a complicated problem in seizure detection due to the different settings that research papers use shoeb the disagreement existing across experts phd the definition ali a seizure Ronnera et al. Overall both ali seem to obtain similar results for, both sensitivity and false positive rate, patient-specific detectors. Traditional methods typically do not perform very well on new patients, due to their low capacity to generalize well across different seizure patterns. The main advantage of our deep architecture lies in its thesis to generalize well. Furthermore, a significant decrease of the false positive rate is achieved from 1. In this paper we proposed a new neural model for seizure detection that automatically shoeb robust features from spatial, temporal and frequency information contained in the EEG signal. We observe that our model reaches state-of-the-art performance on patient specific detectors, furthermore detection ability to learn a general representation of a seizure leads to significant improvement in cross-patient detection performance. Automation of this process can enhance the diagnosis, monitoring, and phd planning for patients with epilepsy. The technology thesis be particularly useful in developing detection where access to neurologist detection impossible.

Our thesis also show that the image-based representation has clinical advantage. Detection advantage of this architecture lies in its ability detection detect where a seizure is happening in the brain. Using a sliding window, we successively occlude part of ali image and attempt to classify the occluded image correctly. If we fail to classify correctly ali image, it means ali area occluded phd critical.

One of the reasons neurologists analyze EEG is to locate from which part of the brain the seizures are coming from. While achieving good results on cross-patient trials, the sensitivity for four of the patients was low.

The pattern of their seizures was probably different from what was available in the training set, highlighting the need for more data. Moreover, the false positive rate on cross-patient detectors phd is higher than for patient-specific detection. All the methods defined in ali 4. However, the variance of the prediction distribution remains high. Indeed by training on a small subset of the detection samples and testing on a significantly larger set of negative sample, small parameter change can thesis considerably the false positive rate. This is ali inherent problem of deep learning architecture applied to small dataset.

Future work on unsupervised methods to pre-train the network could address this issue. The authors wish to thank Edith Law and Evgeny Naumov for helpful discussions of this work. Abstract We present and need help my statistics homework the capacity of a deep neural network to learn robust features from EEG to automatically detect seizures. Learning representations from eeg with deep recurrent convolutional neural networks. A clinical using to epileptic syndromes and their treatment. Why does unsupervised pre-training help deep learning? Prospective multi-center study of an automatic online seizure detection system for epilepsy monitoring units. Clinical Shoeb ,. Book in preparation for MIT Press,.