Summary of "Day 2 (Part 1) The BCI & Neurotechnology Spring School 2026"
Main ideas and lessons conveyed
1) Spring School overview (BCI & neurotechnology, Day 2 Part 1)
- The session begins with a live check-in of remote attendees across multiple countries and welcomes participants to Day 2 (Part 1).
- The Spring School is a 10-day free global event focused on learning and community-building around brain-computer interfaces (BCIs) and neurotechnology.
- It operates as a worldwide network:
- ~140 nations collaborating on brain-related exploration
- Last year: ~90,000 participants from 140 countries
- Participation and infrastructure include:
- Official local hosts (57 mentioned) across many countries
- Live viewing hubs (20) where groups can watch and run local demos/discussions
- Hackathon hosts (16 locations mentioned), e.g. Simon Fraser University, Queen’s University, Long Beach State University
- Streaming options:
- Zoom (chat/Q&A)
- YouTube Live (replay available for ~2 weeks on the Spring School page)
- Additional regional streams in parts of China/Southeast Asia
- Incentives and networking:
- A “unicorn education kit” lottery for the lab/university with the most participants (requires sharing participation list via an Excel sheet request)
- Extra rewards for social media engagement (best/funniest watch videos; LinkedIn posts with coupons)
- Encouragement to join:
- GTE Discord
- a WhatsApp channel for news/insights
- Certification:
- Certificates require attending all sessions/keynotes (plus hackathon presentations on Sunday)
- Not required: participating in the hackathon itself
- An end-of-spring multiple-choice exam (~20–25 questions, ~20 minutes) produces a score usable for university credits
2) Day 2 theme: Invasive BCIs & network-guided neuromodulation
The day focuses on:
- Invasive brain-computer interfaces
- Network-guided neuromodulation
- How to map brain networks and use electrophysiology signals to guide:
- surgical therapy
- closed-loop stimulation
Session content: Main lectures/demos and what they teach
A) Kristoff Capella — High gamma activity & cortico-cortical evoked potentials (CCPs) for network mapping
Core goal
Use electrophysiological signals to:
- localize functional brain areas (e.g., language, motor)
- measure how areas are connected (functional networks)
- preserve critical functions during surgery
Key concepts
- High gamma activation from intracranial recordings as a marker of task-related activity
- Cortico-cortical evoked potentials (CCPs):
- generated via single-pulse electrical stimulation
- produce stereotyped components, often described as:
- early marker: N1 (connectivity marker)
- later component: N2
- Speech arrest / transient aphasia during symptomatic stimulation:
- treated as a gold standard for intraoperative decision-making
- used to stop resection to avoid deficits
Demo-driven workflow (inferred steps)
- Task-based high-gamma mapping
- observe high-gamma during language or motor tasks
- use electrode locations to label functional areas
- Symptomatic stimulation
- stimulate candidate language/motor areas
- observe symptoms (e.g., inability to speak specific words)
- use results to guide resection boundaries
- Single-pulse stimulation for CCPs
- use single pulses (not trains)
- analyze evoked response timing and components (N1/N2) for inferred connectivity
- Integrate “what you stimulate” + “what responds”
- stimulating a known language site and detecting evoked responses elsewhere supports network structure inference
- Broaden mapping options
- ECoG/high gamma, fMRI, TMS, and sedated paradigms for hemisphere dominance
Electrical stimulation safety and parametering
Key points include:
- distinguishes chronic vs short train safety limits
- emphasizes charge density and charge limits, including:
- chronic stimulation: keep below ~50 μC/cm² per phase (as stated)
- short trains: up to ~300 μC/cm² (short range) and an upper stated limit of ~375 μC/cm²
- absolute maximum charge: ~15 μC (as stated)
- uses biphasic stimulation with zero net charge balance
- notes damage risk qualitatively:
-
10 mA/mm² can cause damage
- ~ 20 mA/mm² can burn (with used currents staying below these thresholds)
-
- operational procedure:
- impedance checks before stimulation
- over-impedance triggers “high impedance event”
- compliance voltage limit: ~80 V, with stimulation stopped if exceeded
- demonstrates electrode switching programmatically and corresponding unit effects
Closed-loop / advanced demo within the talk
- software configuration of CCP measurement including:
- stimulation frequency (example: ~1.1 Hz to manage mains interference)
- bifasic waveform choices
- referencing selections
- number of trials (e.g., 33 trials)
- attention to DC-coupled amplifiers and visibility of responses immediately after pulses
Network connectivity quantification using structure
- combines functional network mapping with structural tools:
- diffusion tractography / DTI
- atlases such as FreeSurfer / DRaCULA
- relates CCP component properties (e.g., N1) to:
- tract length
- fiber pathways (e.g., arcuate fasciculus, uncinate fasciculus, cingulum)
- response latency/amplitude vs distance
Seizure network and “disconnection surgery” concept
- applies CCP ideas to epilepsy spread:
- stimulate near seizure onset and track evoked connectivity
- example described: inferred corpus callosum disconnection via disappearance of evoked responses after fiber severing
Concluding message
- CCPs enable functional connectivity mapping tied to stimulation-based labels and can support surgical decisions (e.g., language preservation; epilepsy network surgery).
B) Simony Hem — Probabilistic mapping for group analysis & deep brain stimulation (DBS)
Core goal
Improve DBS targeting and programming using data-driven probabilistic group analysis.
Challenges addressed
- pre-surgical planning: identifying the true therapeutic target
- DBS programming complexity:
- many electrode designs (including directional leads)
- many programming choices
- need decision support learned from patient cohorts
Core methodology
- use volume of tissue activated (VTA) simulations
- steps:
- patient-specific modeling
- segment MRI into tissue types (gray matter, white matter, CSF, blood)
- use post-op imaging to place electrodes
- optionally use DTI anisotropy
- run FEM-based electric field simulations
- map patients into a common template space
- build disease-specific templates using MNI registration + nonlinear registration
- project anatomical structures/atlases into template space
- build probabilistic stimulation maps
- associate each projected VTA with clinical outcome (e.g., tremor reduction)
- aggregate with statistical methods across patients and trials
- patient-specific modeling
Statistical/algorithmic choices explored
- variations in output construction and evaluation:
- number of simulations per voxel
- weighted mean maps (weighting by electric field magnitude and/or amplitude preferences)
- selecting voxels for significant correlation with symptom improvement
- parameter choices explored:
- screening vs best-contact datasets
- number of stimulation positions per patient
- weighting/non-binary activation functions
- choice of statistical tests (including Bayesian t-test)
- sample size stability evaluated via dice coefficient
- reported findings:
- similar behavior across PD and ET cohorts
- more data improves stability of “sweet spot” maps
Clinical utility examples
- identifying therapeutic sweet spots and side-effect regions
- capturing hemispheric asymmetry (not just left-right mirroring)
- testing whether intraoperative sweet spots predict post-op programming outcomes
- predictive modeling:
- train models on VTA features to predict side effects vs therapeutic improvements
- outputs described as multi-class ranges (e.g., <50% vs ≥50% improvement, plus side effects)
Takeaways
- group analysis increases statistical power and supports targeting/programming
- limitation: trade-off between group patterns and patient-specific nuance; normalization errors matter
- future direction: predictive maps to automatically propose individualized contacts/amplitudes and leverage directional lead designs
C) Frans (François) — Electrical stimulation, CCP-like potentials, and intraoperative electrodiagnostic mapping
Core aim
Measure and interpret stimulation-evoked potentials during surgery to infer connectivity and tissue properties in real time.
Key concepts
- compares categories of potentials:
- direct cortico-cortical response (DC/EC-like)
- axon-cortical potential (ACP-like)
- cortico-cortical potential (CCP/CEP-like; connectivity probing)
- emphasizes:
- macro vs micro stimulation (selectivity limitations when not sufficiently focal)
- need to characterize the full measurement chain and handle noise
Signal processing themes
- mitigate strong power-line noise and stimulation artifacts:
- detect reproducible 50 Hz noise
- create a virtual noise reference for subtraction
- blank stimulation artifact repeatedly
- filter using specified bandpass choices
- average while preserving variability across epochs
Interpretation
- early waveform components: linked to synchrony and direct pathway activation
- later components: linked to broader cortical network/inhibitory dynamics
- “distance divergence effect”:
- latency/amplitude/time course shifts for long-range pathways due to conduction differences and summation
Open research problems
- source separation:
- separating true potentials from volume-conduction artifacts in complex surgical environments
- improving bipolar probe selectivity:
- suggests exploring multipolar stimulation or waveform changes
Concluding idea
Intraoperative CCP/ACP-like measurements can support electrodiagnostic connectivity mapping, but require methodology standardization and improved source separation.
D) Patrick Kritner — Real-time closed-loop neuromodulation system (software + hardware + latency control)
Core goal
Build a closed-loop neuromodulation pipeline that responds to brain features with ~10–20 ms latency.
Closed-loop principle
- Closed-loop = sense brain state → compute features → decide → trigger stimulation
- contrasts with open-loop methods that stimulate at fixed times/rhythms without state feedback
System architecture (detailed)
- Signal acquisition
- intracranial electrodes (ECoG/stereo-EEG) or simulated inputs
- example amplifier: Gtech GHighamp
- USB to a PC
- Real-time processing / feature extraction
- software: GTE tools / CQ (with mention of MATLAB/Simulink)
- example features:
- phase/power in frequency bands (alpha/theta/beta/gamma)
- HFOs/spikes
- Decision logic
- thresholds or state-based rules for stimulation triggering
- safety constraints:
- refractory period to prevent oversampling / excessive stimulation rate
- artifact rejection (e.g., muscle motion)
- Stimulation triggering
- stimulator controlled via TTL/digital outputs
- switching unit routes pulses to selected electrode channels
- Feedback evaluation
- measure loop latency and jitter online using TTL-to-TTL timing
- confirm ability to hit phase timing targets
Live demos and what they demonstrate
- Demo 1: Loop latency measurement
- Simulink TTL → amplifier → PC/software TTL detection → output TTL
- reported ~10 ms delay (approximate stated range)
- Demo 2: Alpha phase prediction for closed-loop TMS (conceptual)
- input: 10 Hz alpha sine
- acquire at 1200 Hz
- predict phase ahead to compensate hardware delay
- trigger at ~90° (π/2) phase point
- evaluate onset phase distribution via polar histogram (with some jitter)
- Demo 3: Closed-loop stimulation using a real EEG cap
- detect alpha threshold
- phase prediction triggers stimulation at predicted timing
- Demo 4: Closed-loop stimulation using a switching unit
- route stimulation to the appropriate channel pair in real time
- demonstrate online changeability and switching artifacts
Challenges covered
- latency bottlenecks:
- transfer to acquisition PC + preprocessing + feature extraction + trigger timing
- amplifier latency also contributes
- artifact handling
- robustness via safeguards (e.g., refractory time)
E) Ricky Matsumoto — CEP pioneers lecture: cortico-cortical evoked potentials as network neurosurgery tools
Core idea
CEP/CCP supports effective connectivity mapping between cortical regions by inferring directional connectivity, using stimulation + recording network “signatures” (e.g., N1, N2, waveform morphology).
Development history
- pioneered/advanced concept of using single-pulse cortical stimulation and recording evoked responses in connected cortex via implanted electrodes
- supports effective connectivity framing beyond structural mapping
Clinical usage
- during epilepsy and tumor surgeries:
- identify language network nodes by stimulating candidate sites and recording CEPs elsewhere
- use CEP as an electrophysiological marker of tract integrity (notably the arcuate fasciculus)
- feasibility under awake and general anesthesia conditions (as reported)
Methodological points
- stimulation paradigm:
- single pulses at low frequency (~≤1 Hz)
- multiple trials (e.g., 10–30 trials per site)
- filtering caution:
- avoid filtering choices that remove slower components (warning against certain high-pass settings)
- waveform structure:
- early negative component N1
- later broad component N2
- morphology matters:
- introduces a framework for CEP waveform “canonical” classification beyond just N1 presence/absence
- types cluster by N1/N2 timing/polarity (canonical vs atypical)
- spatial distribution of types may encode anatomical/circuit differences
Network neurosurgery examples
- Arcuate fasciculus integrity
- N1 amplitude/latency shifts after tumor resection
- suggested rule-of-thumb risk marker:
- roughly ~50% N1 amplitude decrease (noting need for larger cohort validation)
- pathology and outcomes:
- tumor type/grade affects latency/amplitude patterns
- CEP monitoring correlates with language outcomes (published/ongoing studies)
Research frontiers
- distance/time effects and long-range potentials
- “real-time connectivity maps” with large normative datasets and automated workflows
- future: adapt CEP logic to probe more axonal/potential types and complex pathways
F) Mayo Clinic / Nuri (Nuri Kapi? in transcript as “Nuri”) — Neural biomarkers for neuromodulation (Parkinson + epilepsy), plus AI for seizure localization
Core goal
Use neural biomarkers to improve:
- DBS programming for Parkinson disease
- early seizure localization for epilepsy
- adaptive/closed-loop therapies
Parkinson disease: DBS biomarkers & adaptive logic
Beta activity as a biomarker
- excessive beta in ~13–30 Hz (or 8–30 Hz) range:
- disappears with levodopa
- disappears with effective stimulation
- returns when stimulation stops
- practical programming hypothesis:
- contact pair with the largest beta often matches clinicians’ best programming choices
- beta could guide automated/closed-loop programming
Long-term robustness
- beta patterns remain recordable 3–7 years after implantation
- measured from externalized leads during battery replacement
Medication + stimulation interaction
- during medication on/off cycles:
- beta suppresses when medication turns on
- HFOs appear around 300–350 Hz
- HFOs may show cross-frequency coupling with beta
- clustering by cross-frequency coupling patterns:
- different clusters respond differently to medication
- objective motor measures correlate better than nurse subjective assessments
Intraoperative AI for DBS lead placement
- AI uses local field potentials during electrode placement trajectories to predict neurosurgeon decisions
- reported performance:
- beta-only predicts decisions ~72%
- beta + HFOs predicts up to ~80%
- disagreement cases often relate to malposition risk (possibly correctable)
Mechanistic exploration of therapeutic frequency
- DBS frequency effects:
- <100 Hz can worsen symptoms
- 100–200 Hz suppresses tremor
-
200 Hz symptoms may return
- explores stimulation-evoked resonant activity / HFO emergence in STN and specificity
Epilepsy: HFOs and AI-based seizure-zone detection
HFO definitions
- ripple: ~80–200 Hz
- fast ripple: ~200–250 Hz up to 500 Hz
Challenges
- artifacts and spikes can mimic HFOs
- healthy cortex can generate physiological HFOs
AI approaches
- unsupervised clustering to separate artifacts from true HFOs
- stereotyped waveform detection to distinguish pathological vs physiological HFOs
- sparse signal processing to detect “signal beauty” vs artifact “ugliness”
- residual-learning denoising:
- intraoperative data is noisy; AI learns waveform-shape dictionaries and denoises accordingly
Category
Educational
Share this summary
Is the summary off?
If you think the summary is inaccurate, you can reprocess it with the latest model.
Preparing reprocess...