Abstracts

Rationale: The NeuroPace Responsive Neurostimulation (RNS) system is a recently FDA-approved closed-loop system, comprised of an implantable microprocessor with intracranial sensors capable of responsive detection and neural stimulation. This device is one of the few treatments available for patients who are ineligible for traditional surgical resection. It is becoming an important part of clinical practice, despite a limited understanding of optimal, patient-specific detection and stimulation parameters. The use of RNS has not been well described outside of the pivotal trials. Additional barriers to advancing our understanding of both acute and chronic effects of RNS are the need to acquire data from NeuroPace and the computational infrastructure to parse, store, and process the data. Here, we describe the establishment of an infrastructure for obtaining, storing, and analyzing RNS data. Methods: Obtaining raw intracranial electroencephalography (iEEG) data from the RNS device entails that NeuroPace extract the data, put it on a flash drive, and mail it to the investigator. Additional device detection and stimulation settings were obtained by screen scraping the NeuroPace Patient Data Management System. We established a novel analytics pipeline to address the problem of understanding data collected by the RNS system and optimizing device parameters for seizure detection and neural stimulation. Results: 12 patients with medically refractory epilepsy who were ineligible candidates for surgical resection have been implanted at our center. Patients ages 22-65 have been implanted for a mean of 72 weeks. A lengthy review of the Business Associate Agreement with NeuroPace was undertaken by the Corporate Compliance Office of the University of Pittsburgh Medical Center. It was determined that data collection could proceed under an existing IRB-approved protocol for our Surgical Epilepsy Brain and Biomarker Databank. The RNS pipeline runs from a dedicated server optimized for big data processing with two 10 core E5 processors, 128 GB RAM and 500 GB solid state work space attached to 16 TB of storage. Automated monthly extracts of iEEG recordings and metadata are loaded into SQL Server, a scalable relational database with built in extract, transform, and load tools that interface with all major analytics platforms. The raw signal data is preprocessed using MATLAB. Deep neural learning using TensorFlow, an open source library developed by Google's Machine Intelligence research organization, assists with seizure detection and characterization. Recordings are reviewed by a team of experts to identify seizures and characterize the onset pattern, which are used to train the algorithms and elucidate features of interest. Finally, we apply time-frequency power analysis to identify biomarkers for seizure type and correlation with stimulation dose. Conclusions: The RNS device offers researchers unprecedented access to years’ worth of chronic human brain activity data. By using human subjects’ data that has never before been available, we hope to improve baseline device settings and programming, which may reduce mean time to seizure reduction, and improve chances for seizure control. This may establish a personalized medicine, machine-learning-based framework for electrographic biomarker discovery that could be used to both tailor stimulation to individual physiology and serve as a model for informing initial programming in new RNS patients. We are exploring options to develop this pipeline as a centralized means for collaborative analysis of RNS data from multiple institutions. Funding: University of Pittsburgh Physician Scientist Training ProgramWalter L. Copeland Fund of The Pittsburgh Foundation