About Us

Real-time neural signal processing, from sensor to clinical insight

Our Mission

BitBlend exists to make neural signal processing faster, more precise, and clinically actionable. We build EEG acquisition and analysis systems that operate in real time, not as a feature, but as a fundamental design constraint that shapes every decision from hardware selection to algorithm architecture.

We collaborate with university medical centers and research institutions across the Netherlands to develop signal processing pipelines that extract meaningful biomarkers from EEG data. Our work spans cross-frequency coupling analysis for early cognitive decline detection, real-time spectral decomposition for neurofeedback, and embedded machine learning classifiers for automated seizure detection.

Our Foundation: Clinical neuroscience demands measurement systems where signal integrity, processing latency, and diagnostic accuracy are inseparable concerns. We engineer all three as a single system.

Our Approach

BitBlend: where digital precision meets neural complexity. We don't use off-the-shelf toolboxes designed for offline research. We build custom signal processing from the ground up in C/C++, optimized for the specific clinical application: the exact frequency bands, the exact latency budget, the exact hardware target.

This matters because EEG signals are measured in microvolts, buried in noise orders of magnitude larger. Extracting clinically relevant features, phase-amplitude coupling, spectral power ratios, event-related potentials, requires algorithms that are not just correct, but fast enough to drive real-time feedback and robust enough to handle the variability of clinical environments.

Our systems close the loop between measurement and intervention. When a neurofeedback protocol requires the brain's current state to be classified and fed back within milliseconds, when an early detection algorithm must process 64 channels continuously without dropping a sample, when a classifier must run inference on an embedded target with a fixed memory budget, that is where our engineering makes the difference.

Core Principle: Build measurement systems that are precise enough to detect what matters, fast enough to act on it, and reliable enough to trust in clinical practice.

Our Knowledge and Expertise

Our main R&D domains and technical focus areas

Algorithmic Design & Signal Processing

We build custom EEG signal processing algorithms in C/C++ rather than relying on research toolboxes like EEGLAB or MNE-Python. This gives us the control needed for clinical deployment: deterministic execution, optimized memory usage, and processing speeds that enable real-time feedback.

What We Do:

Cross-frequency coupling analysis: extracting phase-amplitude coupling biomarkers that detect cognitive decline before structural changes appear
Spectral decomposition using Welch's method and multitaper estimation, optimized for noisy clinical environments
ICA-based artifact removal that separates eye blinks and muscle activity from neural signals without destroying underlying brain data
SIMD-optimized filter implementations that process multiple channels simultaneously
Artifact Subspace Reconstruction for real-time artifact rejection without discarding data

Why It Matters: Off-the-shelf tools are designed for offline research. Clinical systems need algorithms that run continuously, in real time, with guaranteed timing, and that is what we build.

More Info

Hardware Integration & Real-Time Systems

Real-time EEG processing is fundamentally different from offline analysis. In closed-loop neurofeedback, only the worst-case latency matters, not the average. We design acquisition and processing hardware that guarantees deterministic timing from electrode to feedback output.

Our Capabilities:

Integration with clinical-grade ADCs (ADS1299: 24-bit resolution, 1µVpp noise, -120dB CMRR)
DMA-based zero-copy acquisition: samples flow from ADC to processing pipeline without CPU intervention
FPGA-accelerated parallel processing for 64+ channel EEG with pipeline architectures
Closed-loop feedback systems targeting under 10ms total cycle time
Real-time impedance monitoring and signal quality assessment
Bare-metal and RTOS implementations for deterministic execution

Why It Matters: Typical neurofeedback systems operate at 300-1000ms delay. EEG microstates last only 80-120ms. To meaningfully interact with brain dynamics, your system has to be faster than the brain state you're measuring.

More Info

Neuroscience-Driven Technology

Neurofeedback works through operant conditioning: the brain learns to modulate its own oscillations when given real-time feedback. We implement the protocols that make this possible, from SMR training for epilepsy to theta/beta ratio training for ADHD, and we build the signal processing that makes them responsive enough to be effective.

What We Build:

Real-time cognitive state detection from EEG, attention, cognitive load, drowsiness, using spectral power ratios
Event-related potential (ERP) extraction: P300, N400, and mismatch negativity for objective cognitive assessment
Brain-computer interface paradigms: motor imagery classification, P300 spellers, SSVEP-based control
Electrode configuration optimized for user comfort, supporting wet, dry, and ear-EEG systems
Psychometric validation of EEG-based assessments against established clinical instruments

Why It Matters: The effectiveness of neurofeedback depends directly on how quickly and accurately the system can classify brain state and deliver feedback. Our closed-loop architectures make the difference between a protocol that works and one that doesn't.

More Info

Machine Learning for Neural Signals

Standard deep learning architectures fail on EEG data. Datasets are small (100-500 trials), subject variability is enormous, and clinical deployment demands explainability. We use specialized architectures designed specifically for the constraints of neural signal classification.

Our Approach:

EEGNet and compact CNN architectures (2K-10K parameters) whose learned filters match classical neuroscience frequency bands
LSTM and Transformer models for temporal pattern detection: capturing cognitive state changes over seconds to minutes
Transfer learning with subject-adaptive fine-tuning to handle individual EEG variability with minimal calibration
8-bit quantized models deployed on microcontrollers (under 1MB flash) with 3-15ms inference latency
Explainability via SHAP and attention map visualization, required for clinical trust and regulatory compliance

Why It Matters: FDA has authorized over 1,400 AI/ML medical devices. EU MDR classifies EEG diagnostic software as minimum Class IIa. We design our models with regulatory compliance built in from the start, not added as an afterthought.

More Info

Impact

Research collaborations and applied neurotechnology projects

UMC Utrecht Collaboration

Advanced neuroimaging research collaboration with University Medical Center Utrecht's Rudolf Magnus Institute for Neuroscience. Developing cutting-edge MRI visualization algorithms using real-time volumetric rendering, 4D flow imaging, and advanced diffusion tensor imaging (DTI) for enhanced clinical diagnostics.

fMRI Visualization Volumetric Rendering DTI Analysis 4D Flow Imaging Real-time Processing

More Info

Optical MRI Volume Simulation

Revolutionary light transmission simulation framework for analyzing and recreating brain activation areas within MRI volumes. Utilizes Monte Carlo photon transport modeling, advanced ray-tracing algorithms, and optical coherence tomography principles to reconstruct neural activation patterns with unprecedented spatial resolution and temporal accuracy.

Monte Carlo Simulation Photon Transport Ray Tracing Optical Coherence Volume Reconstruction

Research in Progress

Engineering Precision for Neural Science

The brain's electrical activity carries information about cognitive health, neurological disease, and mental state, but only if you can measure it precisely enough. EEG signals are measured in microvolts, buried in noise from muscle activity, eye movements, and power lines. Extracting clinically useful features from this signal requires a combination of sophisticated algorithms and purpose-built hardware.

Cross-frequency coupling analysis can detect functional changes associated with early cognitive decline, changes that appear before structural neurodegeneration is visible on any scan. Real-time spectral decomposition enables neurofeedback protocols that train the brain to modulate its own oscillations. Compact neural networks, small enough to run on a microcontroller, can classify seizure patterns continuously without ever missing a beat.

These capabilities don't come from any single technology. They emerge from the integration of hardware design, signal processing, neuroscience, and machine learning, all engineered as a single system with one goal: closing the loop between what the brain does and what the clinician can act on.

We build the systems that turn neural signals into clinical insight, fast enough to matter, precise enough to trust.

24-bit

ADC resolution at 1µVpp noise floor

<10ms

Closed-loop neurofeedback cycle time

3ms

On-device ML inference per EEG frame