Neural NoiseBCI Annual Review — 2005
1 January–31 December 2005
Introduction
2005 was a year of expanding frontiers and mounting complexity in BCI research. The BrainGate pilot clinical trial continued its slow but historically significant progress: Matthew Nagle completed the majority of his 57 documented training sessions at New England Sinai Hospital between July 2004 and April 2005, and the second BrainGate trial participant — a 55-year-old man with cervical spinal cord injury — received his Utah Array implant at the University of Chicago in April 2005. The accumulation of human data from these two participants was generating rich evidence about the behavior of intracortical recordings in injured human brains over months, including the well-documented decline in signal quality that Nagle experienced after approximately 6.5 months. That signal degradation — likely due to gliosis around the electrode tips — became one of the most discussed open problems in the field, spurring increased research into electrode materials, surface coatings, and flexible array designs intended to reduce the brain’s foreign-body response.
On the pre-clinical front, Andrew Schwartz’s group at the University of Pittsburgh was advancing toward the monkey self-feeding experiments that would make headlines in 2008, while Krishna Shenoy’s Stanford lab published key results on premotor cortex plan-activity decoding at high information rates. The third BCI Competition, BCI Competition III, was organized and its data sets distributed to the research community in 2005, again covering ECoG, P300, and motor imagery paradigms, with a notable first inclusion of ECoG data sets. The competition, whose results would be reviewed in IEEE Transactions on Neural Systems and Rehabilitation Engineering in 2006, again attracted nearly 100 submissions and documented progress in machine learning approaches including support vector machines, regularized linear discriminants, and Bayesian decoding methods applied to neural signals.
A critical development in neuromodulation came with the expanded clinical use of DBS for dystonia under the 2003 HDE, and the growing interest in using local field potentials (LFPs) recorded through DBS electrode contacts as potential BCI control signals — a line of research that would bridge neuromodulation and communication applications. Transcranial magnetic stimulation (TMS) was increasingly used as a research tool to map motor cortex representations, directly informing the interpretation of BCI recording sites, while tDCS (transcranial direct current stimulation) was being explored as a potential tool to enhance motor learning and BCI training, though most work remained in early feasibility stages. The overall picture for 2005 was a field expanding its clinical footprint, encountering the messy realities of long-term neural recording stability in humans, and responding with broader research investment in both electrode technology and neural decoding methodology.
The BrainGate pilot trial also expanded its site infrastructure in 2005: Leigh Hochberg, as the principal investigator at Spaulding Rehabilitation Hospital (Harvard Medical School), was actively recruiting additional trial participants under the FDA-approved protocol, which allowed for up to five human subjects. Cyberkinetics simultaneously began work on a second clinical application of the BrainGate platform — a trial in patients with ALS at Massachusetts General Hospital — recognizing that the technology’s most immediate clinical value might be in providing communication for patients with progressive neurodegenerative disease rather than SCI alone. These parallel efforts required new thinking about the appropriate patient populations, outcome measures, and performance benchmarks for intracortical BCIs in humans.
Timelines
January–March. Matthew Nagle continued his BrainGate training sessions into the new year at New England Sinai Hospital, Stoughton, Massachusetts. His training sessions (57 total, July 2004–April 2005) were demonstrating that he could operate cursor-based computer tasks with notable proficiency for someone using only motor cortex neural signals from a paralyzed brain. Shenoy’s Stanford group was completing analysis of their high-information-rate premotor cortex decoding experiments, finding that brief “plan activity” epochs during a target delay period encoded more information than previously appreciated and could support communication rates that would make cursor-based typing practical. The BCI Competition III data sets were being assembled by the organizing consortium — Graz (Pfurtscheller), Wadsworth (Wolpaw), Birbaumer’s group at Tübingen, and the Berlin BCI group (Blankertz/Müller) — for release to the community.
April–June. The second BrainGate trial participant received his intracortical implant at the University of Chicago in April 2005, with neurosurgeon Richard Penn and Cyberkinetics technicians. Signal quality was initially compromised by electrode contact issues, but was later repaired, allowing the participant to achieve cursor control from months seven through ten of his trial. The BCI Competition III data sets were released to the research community, including for the first time a data set based on ECoG motor imagery from Birbaumer’s group, and P300 speller data from Wadsworth. The Graz group published multiple papers on motor imagery classification, including four-class imagery (left hand, right hand, foot, tongue), expanding the command vocabulary available from a single EEG-based BCI session.
July–September. Nagle’s BrainGate implant was surgically removed on October 18, 2005, at Rhode Island Hospital — not due to a safety concern but because Nagle wanted to undergo placement of a phrenic nerve pacemaker (to reduce dependence on his ventilator), which was incompatible with ongoing BrainGate trial participation. His total implant duration was approximately 16 months. The rich dataset accumulated over his 57 sessions was now available for full analysis, which would yield the landmark Nature paper the following year. Schwartz’s group at Pittsburgh was conducting increasingly sophisticated monkey arm-control experiments, with monkeys learning to use a four-degree-of-freedom robot arm (shoulder, elbow, gripper) for reaching and grasping. The Society for Neuroscience annual meeting in November 2005, held in Washington, D.C., featured a large BCI program, with multiple posters and symposia from Donoghue, Shenoy, Schwartz, Nicolelis, Andersen, and the EEG-BCI community.
October–December. The removal of Nagle’s BrainGate implant on October 18, 2005, closed the first chapter of the clinical trial. Cyberkinetics continued to recruit for additional participants, and Hochberg’s site at Spaulding was actively screening candidates. Shenoy’s Stanford group was finalizing their Nature paper (Santhanam, Ryu, Yu, Afshar, Shenoy) on 6.5 bits/second information transfer from premotor cortex plan activity, which would appear in mid-2006. BCI Competition III submissions closed in late 2005, and the analysis of 99 submissions was underway. The use of support vector machines and ensemble methods showed clear advantages over simple linear discriminants on some data sets, particularly for fine-grained ECoG motor imagery classification. The first publications describing DBS field potential recordings as potential BCI control signals for patients with locked-in syndrome appeared in European neurology journals.
Trends
Long-Term Recording Stability: The Field’s Central Engineering Challenge
The most pressing technical issue revealed by the Nagle trial was the degradation of neural signal quality over time. While Nagle’s BCI performance was robust through the first several months post-implant, neural signals declined significantly by month 6–7, attributed primarily to the brain’s foreign-body response to the rigid silicon Utah Array. Astrocytic encapsulation and neuronal retreat from electrode tips reduced both the number of active recording channels and the signal-to-noise ratio of remaining units. This observation, consistent with what several primate laboratories had documented but more visible in the clinical context, immediately elevated the priority of electrode biocompatibility research. Groups at University of Michigan, University of Utah, and MIT were exploring flexible polymer-based electrode materials, nano-structuring of electrode surfaces, and anti-inflammatory coatings as strategies to prolong stable recording. The issue would remain a defining constraint on intracortical BCIs for the next decade.
High-Rate Communication via Plan-Activity Decoding
Shenoy’s group at Stanford demonstrated that premotor cortex “plan activity” — the sustained neural firing that occurs during the preparatory delay period before a reaching movement — encoded target location with high information content even when sampled over periods as short as 300–600 ms. This was a significant departure from the dominant approach of decoding continuous movement kinematics, and it implied that a BCI optimized for a rapid sequence of discrete target selections (e.g., a letter-by-letter communication prosthesis) could achieve much higher throughput than a continuous trajectory decoder. The Stanford lab estimated information transfer rates up to 6.5 bits per second in their primate experiments, which would correspond to approximately 15 words per minute in a spelling application — the first time the field had a concrete estimate competitive with existing assistive technology alternatives.
BCI Competition III and the Rise of Machine Learning
The BCI Competition III, with its inclusion of ECoG data and continuation of motor imagery and P300 paradigms, further stimulated algorithmic innovation in the field. Regularized classifiers — ridge regression, elastic net, regularized linear discriminant analysis — showed particular advantages when training data was limited, a common constraint in clinical BCI sessions. Support vector machines with radial basis kernels achieved strong performance on ECoG data. Ensemble and committee methods improved P300 detection. Bayesian filtering approaches, particularly particle filters and Kalman filters, were proving their utility for continuous decoding of intended limb trajectories in offline analyses. The Kalman filter, in particular, would become standard in subsequent years for its computational efficiency and natural handling of temporal dependencies in neural spike trains.
ECoG BCI: From Epilepsy Monitoring to Research Platform
The Leuthardt et al. 2004 paper had established that ECoG could support rapid BCI control. In 2005, the Seattle group (Rao, Ojemann, and colleagues at the University of Washington) published additional ECoG-BCI results confirming and extending these findings across a larger patient series (ten subjects), reporting that all patients achieving closed-loop trials attained final target accuracies of 73–100%. These results, published in IEEE Trans. Neural Systems Rehabilitation Engineering in 2006 but based on 2004–2005 data, established that the ECoG approach was reproducible. The growing realization that the epilepsy monitoring surgical procedure — already performed tens of thousands of times annually for pre-surgical evaluation — provided temporary access to cortical signals suitable for BCI research was creating an unusual research opportunity: a population of patients willing and able to participate in short BCI experiments during their monitoring stay, with far lower risk than permanent implants.
ALS Communication: BCI2000 in the Clinic
The Wadsworth Center’s P300 speller, running under BCI2000, was by 2005 being actively tested with ALS patients who had lost all voluntary motor control. Reports from multiple centers indicated that ALS patients who retained the ability to generate P300 responses — which is preserved as long as the patient can attend to a visual stimulus — could achieve communication rates sufficient for practical use: several characters per minute with acceptable accuracy. One well-documented ALS patient reported preferring the BCI speller over his existing eye-gaze system for certain tasks, and was using it for four to six hours per day. These early clinical user reports provided qualitative evidence that non-invasive EEG-BCIs could improve quality of life for patients in the most severe stages of motor neuron disease.
Suggested Titles
- The First Year After: Signal Decay, New Patients, and the Realities of Clinical BCI
- Six-Point-Five Bits per Second: Stanford, Plan Activity, and the Information-Rate Race
- ECoG’s Coming Out: From Epilepsy Monitoring Room to BCI Research Platform
- Matt Nagle’s Legacy: What Sixteen Months of BrainGate Data Revealed
- Stable Signals or Scar Tissue: The Foreign-Body Problem Shapes a Field