Closed-loop Decoder Adaptation Research Articles

Online adaptive group-wise sparse Penalized Recursive Exponentially Weighted N-way Partial Least Square for epidural intracranial BCI.

Motor Brain-Computer Interfaces (BCIs) create new communication pathways between the brain and external effectors for patients with severe motor impairments. Control of complex effectors such as robotic arms or exoskeletons is generally based on the real-time decoding of high-resolution neural signals. However, high-dimensional and noisy brain signals pose challenges, such as limitations in the generalization ability of the decoding model and increased computational demands. The use of sparse decoders may offer a way to address these challenges. A sparsity-promoting penalization is a common approach to obtaining a sparse solution. BCI features are naturally structured and grouped according to spatial (electrodes), frequency, and temporal dimensions. Applying group-wise sparsity, where the coefficients of a group are set to zero simultaneously, has the potential to decrease computational time and memory usage, as well as simplify data transfer. Additionally, online closed-loop decoder adaptation (CLDA) is known to be an efficient procedure for BCI decoder training, taking into account neuronal feedback. In this study, we propose a new algorithm for online closed-loop training of group-wise sparse multilinear decoders using L p -Penalized Recursive Exponentially Weighted N-way Partial Least Square (PREW-NPLS). Three types of sparsity-promoting penalization were explored using L p with p = 0., 0.5, and 1. The algorithms were tested offline in a pseudo-online manner for features grouped by spatial dimension. A comparison study was conducted using an epidural ECoG dataset recorded from a tetraplegic individual during long-term BCI experiments for controlling a virtual avatar (left/right-hand 3D translation). Novel algorithms showed comparable or better decoding performance than conventional REW-NPLS, which was achieved with sparse models. The proposed algorithms are compatible with real-time CLDA. The proposed algorithm demonstrated good performance while drastically reducing the computational load and the memory consumption. However, the current study is limited to offline computation on data recorded with a single patient, with penalization restricted to the spatial domain only.

Plug-and-play control of a brain–computer interface through neural map stabilization

Brain-computer interfaces (BCIs) enable control of assistive devices in individuals with severe motor impairments. A limitation of BCIs that has hindered real-world adoption is poor long-term reliability and lengthy daily recalibration times. To develop methods that allow stable performance without recalibration, we used a 128-channel chronic electrocorticography (ECoG) implant in a paralyzed individual, which allowed stable monitoring of signals. We show that long-term closed-loop decoder adaptation, in which decoder weights are carried across sessions over multiple days, results in consolidation of a neural map and 'plug-and-play' control. In contrast, daily reinitialization led to degradation of performance with variable relearning. Consolidation also allowed the addition of control features over days, that is, long-term stacking of dimensions. Our results offer an approach for reliable, stable BCI control by leveraging the stability of ECoG interfaces and neural plasticity.

Robust Brain-Machine Interface Design Using Optimal Feedback Control Modeling and Adaptive Point Process Filtering.

Much progress has been made in brain-machine interfaces (BMI) using decoders such as Kalman filters and finding their parameters with closed-loop decoder adaptation (CLDA). However, current decoders do not model the spikes directly, and hence may limit the processing time-scale of BMI control and adaptation. Moreover, while specialized CLDA techniques for intention estimation and assisted training exist, a unified and systematic CLDA framework that generalizes across different setups is lacking. Here we develop a novel closed-loop BMI training architecture that allows for processing, control, and adaptation using spike events, enables robust control and extends to various tasks. Moreover, we develop a unified control-theoretic CLDA framework within which intention estimation, assisted training, and adaptation are performed. The architecture incorporates an infinite-horizon optimal feedback-control (OFC) model of the brain’s behavior in closed-loop BMI control, and a point process model of spikes. The OFC model infers the user’s motor intention during CLDA—a process termed intention estimation. OFC is also used to design an autonomous and dynamic assisted training technique. The point process model allows for neural processing, control and decoder adaptation with every spike event and at a faster time-scale than current decoders; it also enables dynamic spike-event-based parameter adaptation unlike current CLDA methods that use batch-based adaptation on much slower adaptation time-scales. We conducted closed-loop experiments in a non-human primate over tens of days to dissociate the effects of these novel CLDA components. The OFC intention estimation improved BMI performance compared with current intention estimation techniques. OFC assisted training allowed the subject to consistently achieve proficient control. Spike-event-based adaptation resulted in faster and more consistent performance convergence compared with batch-based methods, and was robust to parameter initialization. Finally, the architecture extended control to tasks beyond those used for CLDA training. These results have significant implications towards the development of clinically-viable neuroprosthetics.

Optimal calibration of the learning rate in closed-loop adaptive brain-machine interfaces.

Closed-loop decoder adaptation (CLDA) can improve brain-machine interface (BMI) performance. CLDA methods use batches of data to refit the decoder parameters in closed-loop operation. Recently, dynamic state-space algorithms have also been designed to fit the parameters of a point process decoder (PPF). A main design parameter that needs to be selected in any CLDA algorithm is the learning rate, i.e., how fast should the decoder parameters be updated on the basis of new neural observations. So far, the learning rate of CLDA algorithms has been selected empirically using ad-hoc methods. Here we develop a principled framework to calibrate the learning rate in adaptive state-space algorithms. The learning rate introduces a trade-off between the convergence rate and the steady-state error covariance of the estimated decoder parameters. Hence our algorithm first finds an analytical upper-bound on the steady-state error covariance as a function of the learning rate. It then finds the inverse mapping to select the optimal learning rate based on the maximum allowable steady-state error. Using numerical BMI experiments, we show that the calibration algorithm selects the optimal learning rate that meets the requirement on steady-state error level while achieving the fastest convergence rate possible corresponding to this steady-state level.

High-performance brain-machine interface enabled by an adaptive optimal feedback-controlled point process decoder.

Brain-machine interface (BMI) performance has been improved using Kalman filters (KF) combined with closed-loop decoder adaptation (CLDA). CLDA fits the decoder parameters during closed-loop BMI operation based on the neural activity and inferred user velocity intention. These advances have resulted in the recent ReFIT-KF and SmoothBatch-KF decoders. Here we demonstrate high-performance and robust BMI control using a novel closed-loop BMI architecture termed adaptive optimal feedback-controlled (OFC) point process filter (PPF). Adaptive OFC-PPF allows subjects to issue neural commands and receive feedback with every spike event and hence at a faster rate than the KF. Moreover, it adapts the decoder parameters with every spike event in contrast to current CLDA techniques that do so on the time-scale of minutes. Finally, unlike current methods that rotate the decoded velocity vector, adaptive OFC-PPF constructs an infinite-horizon OFC model of the brain to infer velocity intention during adaptation. Preliminary data collected in a monkey suggests that adaptive OFC-PPF improves BMI control. OFC-PPF outperformed SmoothBatch-KF in a self-paced center-out movement task with 8 targets. This improvement was due to both the PPF's increased rate of control and feedback compared with the KF, and to the OFC model suggesting that the OFC better approximates the user's strategy. Also, the spike-by-spike adaptation resulted in faster performance convergence compared to current techniques. Thus adaptive OFC-PPF enabled proficient BMI control in this monkey.

Continuous closed-loop decoder adaptation with a recursive maximum likelihood algorithm allows for rapid performance acquisition in brain-machine interfaces.

Closed-loop decoder adaptation (CLDA) is an emerging paradigm for both improving and maintaining online performance in brain-machine interfaces (BMIs). The time required for initial decoder training and any subsequent decoder recalibrations could be potentially reduced by performing continuous adaptation, in which decoder parameters are updated at every time step during these procedures, rather than waiting to update the decoder at periodic intervals in a more batch-based process. Here, we present recursive maximum likelihood (RML), a CLDA algorithm that performs continuous adaptation of a Kalman filter decoder's parameters. We demonstrate that RML possesses a variety of useful properties and practical algorithmic advantages. First, we show how RML leverages the accuracy of updates based on a batch of data while still adapting parameters on every time step. Second, we illustrate how the RML algorithm is parameterized by a single, intuitive half-life parameter that can be used to adjust the rate of adaptation in real time. Third, we show how even when the number of neural features is very large, RML's memory-efficient recursive update rules can be reformulated to also be computationally fast so that continuous adaptation is still feasible. To test the algorithm in closed-loop experiments, we trained three macaque monkeys to perform a center-out reaching task by using either spiking activity or local field potentials to control a 2D computer cursor. RML achieved higher levels of performance more rapidly in comparison to a previous CLDA algorithm that adapts parameters on a more intermediate timescale. Overall, our results indicate that RML is an effective CLDA algorithm for achieving rapid performance acquisition using continuous adaptation.

Closed-Loop Decoder Adaptation Shapes Neural Plasticity for Skillful Neuroprosthetic Control

Neuroplasticity may play a critical role in developing robust, naturally controlled neuroprostheses. This learning, however, is sensitive to system changes such as the neural activity used for control. The ultimate utility of neuroplasticity in real-world neuroprostheses is thus unclear. Adaptive decoding methods hold promise for improving neuroprosthetic performance in nonstationary systems. Here, we explore the use of decoder adaptation to shape neuroplasticity in two scenarios relevant for real-world neuroprostheses: nonstationary recordings of neural activity and changes in control context. Nonhuman primates learned to control a cursor to perform a reaching task using semistationary neural activity in two contexts: with and without simultaneous arm movements. Decoder adaptation was used to improve initial performance and compensate for changes in neural recordings. We show that beneficial neuroplasticity can occur alongside decoder adaptation, yielding performance improvements, skill retention, and resistance to interference from native motor networks. These results highlight the utility of neuroplasticity for real-world neuroprostheses.

Subject-specific modulation of local field potential spectral power during brain–machine interface control in primates

Objective. Intracortical brain–machine interfaces (BMIs) have predominantly utilized spike activity as the control signal. However, an increasing number of studies have shown the utility of local field potentials (LFPs) for decoding motor related signals. Currently, it is unclear how well different LFP frequencies can serve as features for continuous, closed-loop BMI control. Approach. We demonstrate 2D continuous LFP-based BMI control using closed-loop decoder adaptation, which adapts decoder parameters to subject-specific LFP feature modulations during BMI control. We trained two macaque monkeys to control a 2D cursor in a center-out task by modulating LFP power in the 0–150 Hz range. Main results. While both monkeys attained control, they used different strategies involving different frequency bands. One monkey primarily utilized the low-frequency spectrum (0–80 Hz), which was highly correlated between channels, and obtained proficient performance even with a single channel. In contrast, the other monkey relied more on higher frequencies (80–150 Hz), which were less correlated between channels, and had greater difficulty with control as the number of channels decreased. We then restricted the monkeys to use only various sub-ranges (0–40, 40–80, and 80–150 Hz) of the 0–150 Hz band. Interestingly, although both monkeys performed better with some sub-ranges than others, they were able to achieve BMI control with all sub-ranges after decoder adaptation, demonstrating broad flexibility in the frequencies that could potentially be used for LFP-based BMI control. Significance. Overall, our results demonstrate proficient, continuous BMI control using LFPs and provide insight into the subject-specific spectral patterns of LFP activity modulated during control.

Design and Analysis of Closed-Loop Decoder Adaptation Algorithms for Brain-Machine Interfaces

Closed-loop decoder adaptation (CLDA) is an emerging paradigm for achieving rapid performance improvements in online brain-machine interface (BMI) operation. Designing an effective CLDA algorithm requires making multiple important decisions, including choosing the timescale of adaptation, selecting which decoder parameters to adapt, crafting the corresponding update rules, and designing CLDA parameters. These design choices, combined with the specific settings of CLDA parameters, will directly affect the algorithm's ability to make decoder parameters converge to values that optimize performance. In this article, we present a general framework for the design and analysis of CLDA algorithms and support our results with experimental data of two monkeys performing a BMI task. First, we analyze and compare existing CLDA algorithms to highlight the importance of four critical design elements: the adaptation timescale, selective parameter adaptation, smooth decoder updates, and intuitive CLDA parameters. Second, we introduce mathematical convergence analysis using measures such as mean-squared error and KL divergence as a useful paradigm for evaluating the convergence properties of a prototype CLDA algorithm before experimental testing. By applying these measures to an existing CLDA algorithm, we demonstrate that our convergence analysis is an effective analytical tool that can ultimately inform and improve the design of CLDA algorithms.

Closed-Loop Decoder Adaptation on Intermediate Time-Scales Facilitates Rapid BMI Performance Improvements Independent of Decoder Initialization Conditions

Closed-loop decoder adaptation (CLDA) shows great promise to improve closed-loop brain-machine interface (BMI) performance. Developing adaptation algorithms capable of rapidly improving performance, independent of initial performance, may be crucial for clinical applications where patients have limited movement and sensory abilities due to motor deficits. Given the subject-decoder interactions inherent in closed-loop BMIs, the decoder adaptation time-scale may be of particular importance when initial performance is limited. Here, we present SmoothBatch, a CLDA algorithm which updates decoder parameters on a 1-2 min time-scale using an exponentially weighted sliding average. The algorithm was experimentally tested with one nonhuman primate performing a center-out reaching BMI task. SmoothBatch was seeded four ways with varying offline decoding power: 1) visual observation of a cursor ( n = 20), 2) ipsilateral arm movements ( n = 8), 3) baseline neural activity ( n = 17), and 4) arbitrary weights ( n = 11). SmoothBatch rapidly improved performance regardless of seeding, with performance improvements from 0.018 ±0.133 successes/min to > 8 successes/min within 13.1 ±5.5 min ( n = 56). After decoder adaptation ceased, the subject maintained high performance. Moreover, performance improvements were paralleled by SmoothBatch convergence, suggesting that CLDA involves a co-adaptation process between the subject and the decoder.