Future of seizure prediction and intervention: closing the loop

Abstract

The ultimate goal of epilepsy therapies is to provide seizure control for all patients while eliminating side effects. Improved specificity of intervention through on-demand approaches may overcome many of the limitations of current intervention strategies. This article reviews the progress in seizure prediction and detection, potential new therapies to provide improved specificity, and devices to achieve these ends. Specifically, we discuss (1) potential signal modalities and algorithms for seizure detection and prediction, (2) closed-loop intervention approaches, and (3) hardware for implementing these algorithms and interventions. Seizure prediction and therapies maximize efficacy, whereas minimizing side effects through improved specificity may represent the future of epilepsy treatments.

Figure 1

Hypothetical closed-loop experimental protocol for suppressing seizures using multimodal recordings of physiological activity. From the recordings, features are extracted and a classifier is applied to detect seizure activity, or a pre-seizure state for prediction. Upon detection of an event, the device triggers a therapy.

Figure 2

Classification of EEG with computational models. Depth EEG is collected (left panel). A model of EEG data (middle) is fit to the data through data assimilation. The parameter space of the model is explored to make a landscape that determines the different behaviors that the model can produce as a function of the parameters (right, different colors represent different patterns of activity). By fitting the model to the data using data assimilation tools, the current state of the neural activity can be estimated. By knowing how the current state relates to the parameter landscape of behaviors, it can be determined if the activity is approaching the boarder of pathological behavior and a systematic intervention can be planned.

Figure 3

Types of closed loop controllers. Left, A Gradient Descent Controller adjusts parameters and measures effect, such as seizure frequency. Through small changes on a daily basis the algorithm can traverse a parameter landscape to find an optimal solution. In a Reactive Therapy, middle, signals are processed and when an event is detected in the estimated state of the patient. In this example a stereotyped stimulus is applied; however, statistical tools can be applied to optimize stimulus parameters for reactive therapy. In Adaptive therapy, right, the stimulus is modulated by the state of the patient or can be used to trigger phasic stimulation with millisecond precision.

Figure 4

Optogenetic suppression of limbic seizures. Mice expressing halorhodopsin (a) in excitatory cells were injected unilaterally with kainate (KA) into the hippocampus (b). Inhibition of excitatory cells by light delivery (light gray horizontal bars) following seizure detection (c, dark grey vertical bars) leads to seizure suppression (d) compared with control events (e). Most seizures were reduced to less than one second post-detection duration (f). Gray bars: events receiving light; black hashed bars: no-light internal controls. Reproduced with modification with permission from Krook-Magnuson et al (2013).

Figure 5

Hypothetical wireless bidirectional adaptive multimodal neurostimulator interface. This device would capture superficial and deep physiological signals using a multimodal neural interface (stars and grey arrows in cartoon brain). These signals are digitized and analyzed in real-time using computationally efficient circuit architectures. Intervention is applied using electrical and/or optical stimulation. Signals are wirelessly transmitted via bluetooth to a nearby device or the patient’s smart phone. Data files can be instantly uploaded to a secure cloud-drive, and notifications sent to the clinician.