Date of Award
Doctor of Philosophy (PhD)
Continuous medical data monitoring is playing an increasingly important role in patient care, both in and out the hospital. Diagnosing and treating patients with epilepsy is especially reliant on continuous EEG monitoring to identify and respond to seizures. However, as use of continuous EEG becomes more common, both for long-term inpatient monitoring and in ambulatory or implanted devices, the burden of study interpretation is rapidly outpacing available physician resources. In particular, the advent of implanted neuroresponsive devices for treating medically-refractory epilepsy is generating large, streaming datasets potentially lasting for several years and containing hundreds of seizures. The current need for manual review of long-term, continuous EEG data introduces tremendous health care costs and can result in significant delays in seizure diagnosis and treatment. Automated data processing is essential to improve data usage, accurately and rapidly detect seizures, and provide scalability in clinical practice. This thesis aims to develop platforms for automated data analysis and event detection using custom machine learning algorithms for application in the intensive care unit and in implanted neural devices.
The work presented in this thesis progresses through the development of each component of an automated data analysis platform. The first section describes a system for real-time data analysis and caretaker notification in the ICU, with a focus on the process necessary to harness multi-modal data from clinical recording sources. The next section details the process of developing machine learning algorithms for seizure detection. In this section, I present novel seizure detection strategies as well as a competition designed to crowdsource algorithm development. This work produced several highly-accurate, open-source seizure detection methods, validated in extended human implanted device data, along with pipelines to facilitate algorithm application and benchmarking in new datasets. The last section covers the integration of data management and seizure detection for implementation in next-generation medical devices. I present a novel paradigm to leverage cloud computing resources for seizure detection in an implanted device. This system is then implemented in vivo using a canine epilepsy model, with real-time seizure detection on streaming data from Medtronic’s RC+S neurostimulating device.
These algorithms and flexible analysis platforms are a step toward automating analysis of EEG data for epilepsy therapy. It is my hope that such systems will improve medical data usage, reshape caretaker workflow, and increase the clinical power of continuous medical monitoring.
Baldassano, Steven Nicholas, "Translational Machine Learning For Epilepsy Therapy" (2017). Publicly Accessible Penn Dissertations. 3028.