Download Machine Learning to Optimize Embedded Adaptive Deep Brain Stimulation Book in PDF, ePub and Kindle
This thesis focuses on the development and application of novel machine learning approaches to the problem of optimization in adaptive deep brain stimulation. As brain-computer and brain-machine interfacing has rapidly developed in the past few years, attention in the relevant research has shifted from proof-of-concept to proof-of-feasibility. One of the first and, therefore, best-developed neurotechnologies is deep brain stimulation (DBS). DBS is a surgical intervention prescribed for several treatment-refractory neurological conditions. First, a stimulating electrode is chronically implanted into a condition-specific deep brain structure. The parameters of the stimulation provided by this electrode are then set bya clinician. Stimulation remains at these parameters continuously unless a patient actively chooses to disable their treatment. As has been repeatedly demonstrated, DBS is a safe and effective treatment for a number of movement disorders and is under active investigation for potential use in several psychiatric conditions.DBS, however, is not a panacea. Battery replacement requires a revision surgery, and even rechargeable systems0́9 batteries must generally be replaced at least once within a device's lifetime. Additionally, DBS therapy is associated with a number of unpleasant and sometimes dangerous side effects. These side effects can range from transient paresthesias to episodes of depression and mania, and are broadly correlated with high levels of stimulation over long periods of time. In addition to concerns over battery life and side effects, the programming procedure for DBS is based primarily on a back-and-forth between clinicians and patients in a clinical setting. If we define "optimal" treatment as the most complete suppression of symptoms with the least manifestation of side effects, then achieving the corresponding settings is the goal of this procedure. The time consuming nature of this procedure, when considered in the context of clinical time constraints and patient fatigue, means that the parameters selected are far more likely to be the first passable setting than the truly optimal one. Patients are generally given the ability to disable their stimulation or select from a small range of amplitudes, but cannot actively reprogram their devices outside of a clinical setting. One technique with the potential to alleviate concerns about side effects and battery life is adaptive deep brain stimulation (aDBS). aDBS refers to any method that uses feedback on a patient's state to modulate stimulation parameters in real time. This feedback could come in the form of gyroscope and accelerometer data in the case of movement disorders, or could be derived from neural signals that are correlated with the onset of symptoms. These signals are then processed and meaningful features extracted from them, which may in turn be used to determine the appropriate stimulation parameters. This ensures that stimulation is only applied as needed. It is important to note that aDBS systems also intrinsically expand the state space forDBS programming, potentially adding yet more complexity to an already laborious procedure. As directional leads become more common in DBS devices, this expanded programming state space further reduces the likelihood of optimal settings being reached. A twin requirement to developing effective aDBS systems is thus the design of a streamlined procedure for parameter optimization in DBS programming. This may be accomplished through the introduction of an automated programming pipeline. Through the collection of quantified data on symptom severity through the use of gyroscope or accelerometer data and the digitization of patient feedback on side effects, this pipeline could considerably speed testing.In addition, the digitization of the data required to analyze aDBS parameter performance could be integrated with modern optimization techniques, such that a personalized optimal treatment may be determined to within an increased degree of certainty. This work details approaches for resolving these deeply interwoven problems in aDBS treatment through insights from the fields of machine learning and optimization. We begin by considering the current state of the art in adaptive deep brain stimulation, its accomplishments, but especially its limitations. The principal limitations are: a general reliance on distributed systems that hinder free movement in patients; a focus solely on computationally inexpensive, but potentially suboptimal, binary aDBS control strategies; and the lack of an effective pipeline to deploy optimization methods during or after programming. Furthermore, recognition that these limitations are fundamentally interrelated implies that an integrated approach is required. The three key developments detailed in this work are thus themselves closely interrelated.Despite minor reductions in power savings, fully embedded binary aDBS is specifically de-signed to maximize therapeutic efficacy and ease of programming. Our results demonstrate that such a system is prepared for widespread studies in movement disorders. Our graded aDBS system yielded inconclusive results with regards to power savings and therapeutic efficacy. However, basing our approach to feature selection for symptom estimation from neural data on a model-free foundation has yielded promising evidence for relying on data-driven feature extraction. This approach to feature extraction intrinsically requires less direct programming, and instead maximizes the insights that may be gained from the data itself. Our development of a pipeline for automated programming of DBS parameters based on inertial measurements and patient feedback on side effects was designed to generalize easily into future integration with aDBS programming procedures. Finally, our computational approach to extracting information from a tablet- and mobile-based application demonstrates that semi- or fully-automated remote symptom assessment has the potential to significantly improve the future delivery of optimized, individualized treatment. Throughout this work, integration is a key component of discussion and consideration. Each result extracted from the developments discussed herein represents a small step away from the current standard of care. Considered individually, these steps would be taken incompletely different directions. It is the hope of the author that this work instead constitutes a realignment of these disparate goals and an explicit recognition of their inter-relatedness. Only by treating these problems as different faces of the same die can we arrive at truly optimized personalized treatment in aDBS.