A translational platform for prototyping closed-loop neuromodulation systems

Abstract

While modulating neural activity through stimulation is an effective treatment for neurological diseases such as Parkinson's disease and essential tremor, an opportunity for improving neuromodulation therapy remains in automatically adjusting therapy to continuously optimize patient outcomes. Practical issues associated with achieving this include the paucity of human data related to disease states, poorly validated estimators of patient state, and unknown dynamic mappings of optimal stimulation parameters based on estimated states. To overcome these challenges, we present an investigational platform including: an implanted sensing and stimulation device to collect data and run automated closed-loop algorithms; an external tool to prototype classifier and control-policy algorithms; and real-time telemetry to update the implanted device firmware and monitor its state. The prototyping system was demonstrated in a chronic large animal model studying hippocampal dynamics. We used the platform to find biomarkers of the observed states and transfer functions of different stimulation amplitudes. Data showed that moderate levels of stimulation suppress hippocampal beta activity, while high levels of stimulation produce seizure-like after-discharge activity. The biomarker and transfer function observations were mapped into classifier and control-policy algorithms, which were downloaded to the implanted device to continuously titrate stimulation amplitude for the desired network effect. The platform is designed to be a flexible prototyping tool and could be used to develop improved mechanistic models and automated closed-loop systems for a variety of neurological disorders.

Keywords: automation; closed-loop; hippocampus; neuromodulation; prototyping; seizure.

Figure 1

Simplified model of a closed-loop neuromodulation system.

Figure 2

Generalized framework for a learning agent (reproduced with permission; WikiCommons).

Figure 3

Mapping a generalized learning model to the neuromodulation system; the components that are implanted are highlighted by the dashed box. The shaded boxes represent the implanted components that interface with the environment. Algorithm prototyping occurs in the agent where the physician and researcher can generate new algorithms based on historical data and algorithm performance.

Figure 4

Functional flow diagram of the hybrid implantable system with the internal and external partitions denoted.

Figure 5

Functional flow for data annotation and classification using the external software tool.

Figure 6

Closed-loop neuromodulation system implanted in an ovine model. The figure is reproduced from Stanslaski et al. (2012) with permissions from the IEEE.

Figure 7

After-discharge duration as a function of beta band power increase from suppressed baseline. High amplitude stimulation parameters were kept constant in a given session and were always determined to be sufficient to initiate an AD.

Figure 8

Determination of the hippocampal network transfer function between stimulation and beta band spectral power. There is an initial reduction in beta band power at low stimulation amplitudes, followed by an increase in beta band power at higher stimulation amplitudes, resulting in occasional AD during stimulation at 1.5 V.

Figure 9

Training the classifier to detect the onset of after-discharges in the presence of stimulation. (A) Is a dataset with a representative AD. “Discard transient” periods refer to portions of the signal that were not used in training. (B) Is a dataset for providing stimulation artifact without an AD present. (C) Provides the histogram of detection states versus distance from the classifier boundary. (D) Estimates the true positive (TP) and false positive (FP) percentages based on the classifier. (E) Shows the impact of onset and termination constraint logic on detector specificity by overlaying the estimated detector state with recorded data files.

Figure 10

Hybrid system validation of the auto-shutoff algorithm for preventing sustained after-discharges in the hippocampus.

Figure 11

The embedded control policy for modulating hippocampal network dynamics. Color codes at the top will be used to mark states in the resulting data summary.

Figure 12

Data sample from embedded algorithm (Figure 11). The sample demonstrates data associated with detection of seizure-like events in the presence and absence of stimulation and change stimulation parameters, resulting in no observed after-discharges. “Pre-detection” refers to the period of time when the onset or termination constraint has not yet been met.