Feature Selection and Decoder Design for Closed-loop Neural Interface Learning

Abstract

Neural interfaces, ranging from non-invasive scalp electroencephalography (EEG) and pe-ripheral electromyography (EMG) to invasive intracortical arrays, hold transformative po- tential for restoring motor and communicative independence to individuals with neurological impairments. However, the translation of these systems from controlled laboratory settings to robust, everyday clinical use is hindered by a universal scalability bottleneck. Across all modalities, advances in recording hardware now permit the simultaneous monitoring of hundreds to thousands of channels, creating a “data deluge” that overwhelms current decoding frameworks. This expansion in feature space presents a critical challenge: clinical systems lack the computational efficiency to process high-dimensional feature sets within the strict power and latency constraints of portable or implantable hardware. Furthermore, the inherent non-stationarity of biological signals—whether cortical or peripheral—necessitates adaptive frameworks that can sustain performance over long durations without burdensome recalibration. This dissertation establishes principled computational frameworks to opti- mize feature selection and decoder design, effectively bridging the gap between expanding sensor capabilities and the requirements of real-time, robust control. First, to motivate the need for dimensionality reduction, we characterized the distribu-tion of task information across biophysical, spatial, and spectral scales. Using simultaneous recordings across multiple physiological scales, we demonstrated that neural feature vari- ance is highly redundant and spatially fractured. We identified “hub” electrodes that, while strongly correlated with broad population dynamics, often encode minimal task-relevant information. These findings challenge the standard practice of indiscriminate feature inclu- sion and provide the physiological justification for disentangling feature selection from task decoding. Second, we developed a novel framework for Adaptive Feature Selection to addressthe instability of high-dimensional inputs in closed-loop settings. We demonstrated that applying standard, static feature selection methods directly to online data frequently leads to performance degradation due to statistical volatility. To overcome this, we introduced a dynamic selection algorithm governed by temporal continuity constraints. This approach autonomously identifies and tracks the informative subspace in real-time, stabilizing control performance while significantly reducing the computational load required for decoding. Finally, we addressed the behavioral complexity of real-world use by developing a closed-loop decoder for Hybrid Multitasking. Moving beyond single-degree-of-freedom control, we designed a system that enables users to simultaneously perform continuous tracking and discrete classification. Our results indicate that participants can achieve efficient control over these complex, interfering task demands, and that the rate of user learning is strictly governed by specific decoder adaptation parameters. Collectively, this thesis advances the critical transition of neural interface technologytoward scalable, “plug-and-play” architectures. By solving the dual challenges of efficient online feature selection and robust multitasking control, this work contributes essential methodologies for the development of high-bandwidth clinical systems capable of restoring complex human agency.