Search

Closed-loop brain-implantable neuromodulation devices are a new treatment option for patients with refractory epilepsy. Seizure detection algorithms implemented on such devices are subject to strict power and area constraints. Deep learning methods, though very powerful, tend to have high computational complexity and thus are typically impractical for resource-constrained neuromodulation devices. In this paper, we propose a compact and hardware-efficient one-dimensional convolutional neural network (1D CNN) structure for patient-specific early seizure detection. Feature extraction techniques and a novel initialization method based on the forward-chaining training and testing scheme are used to improve model performance. Our compact model achieves similar accuracy to that of support vector machines, the state-of-the-art method for seizure detection, while consuming over 20x less power.

Epilepsy is a neurological disorder which causes seizures in over 65 million people worldwide. Recently developed implantable therapeutic devices aim to prevent symptoms by applying acute electrical stimulation to the seizure-generating brain region in response to activity detected by on-device machine learning hardware. Many training algorithms require an equal number of examples for each target class (e.g. normal activity and seizures), and performance can suffer if this condition is not satisfied. In the case of epilepsy, poor performance can cause seizures to be missed, or stimulation to be applied erroneously. As there is an abundance of normal (interictal) data in clinical EEG recordings, but seizures are rare events (less than 1% of the dataset), the data available for training is severely imbalanced. There are several conventional pre-processing methods used to address imbalanced class learning, such as down-sampling of the majority class and up-sampling of the minority class, but each have performance drawbacks. This paper presents an improved method which involves reducing the majority class down to the most effective interictal outlier samples. Outliers are determined by using Exponentially Decaying Memory Signal Energy (EDMSE) features with Isolation Forests and an ANOVA-based method, which involves comparing a moving feature window to a baseline reference window. Outlier-based sampling is tested with two classifiers (KNN and Logistic Regression) and achieves higher accuracy (∼2% increase) and fewer false positives (∼38% decrease), along with a lower latency (∼3 seconds shorter) compared to conventional training set pre-processing methods.

Growing evidence suggests that cross-frequency coupling (CFC) is a key mechanism in neuronal computation, communication, and learning in the brain. Abnormal CFC has been implicated in pathological brain states such as epilepsy and Parkinson's disease. A reduction in excessive coupling has been shown in effective neuromodulation treatments, suggesting that CFC may be a useful feedback measure in closed-loop neural stimulation devices. However, processing latency limits the responsiveness of such systems. A VLSI architecture is presented which implements three selectable measures of CFC to enable the application specific trade-off between low-latency and high-accuracy processing. The architecture is demonstrated using in-vitro human neocortical slice recordings, with a latency of 48ms.