How to build a brain–computer interface from beginning to end

Fabien Lotte, Pauline Dreyer, Sébastien Rimbert, David Trocellier, Marc WelterElsevier 2025 · DOI · OpenAlex

Introduction

Building a Brain-Computer Interface (BCI) is a high-stakes, multi-disciplinary puzzle that often overwhelms newcomers. For a novice designer, the BCI landscape is a daunting task characterized by a dizzying array of choices across neuroscience, signal processing, and human-computer interaction. There is no one-size-fits-all solution; a BCI designed for a locked-in patient requires a fundamentally different architecture than one built for a high-speed gaming application.

This paper provides a comprehensive roadmap, guiding designers through the entire lifecycle—from defining user-centric objectives and selecting neural markers to navigating the complex world of machine learning pipelines and neuroethics. The authors aim to move the field away from "trial and error" by providing a structured formalization for BCI design.

The BCI Lifecycle

The paper functions as a meta-framework, breaking the BCI system into a closed-loop interaction involving several technical and non-technical components. This standard pipeline ensures that every stage of the interaction is accounted for, from the user's brain activity to the feedback they receive.

The standard BCI loop: from brain activity to user feedback.

As illustrated in the BCI loop, the process involves:

  • Neuromarkers: The production of specific brain activity patterns.
  • Sensors: The measurement of brain activity (e.g., EEG, ECoG).
  • Signal Processing & Machine Learning: The translation of raw signals into actionable commands.
  • User Interface & Application: The software environment the user interacts with.
  • Feedback & User Training: The mechanism that helps the user improve their control over the system.

Categorization of Paradigms and Machine Learning

To simplify the design process, the authors categorize BCI paradigms into three main types:

  1. Active BCIs: These rely on voluntary tasks, such as motor imagery (e.g., imagining moving a hand).
  2. Reactive BCIs: These utilize responses to external stimuli, such as the P300 wave or Steady-State Visual Evoked Potentials (SSVEP).
  3. Passive BCIs: These monitor involuntary states, such as mental workload or emotional stress, without requiring active control.

In tandem with these paradigms, the authors categorize current machine learning (ML) approaches into a three-tier taxonomy. Raw Classifiers (like LDA or SVM) serve as simple, interpretable baselines. Riemannian Classifiers are highlighted as the current state-of-the-art for EEG, as they operate on covariance matrices to handle high signal variability. Finally, Deep Learning architectures (CNNs, RNNs) are recommended for massive, labeled datasets and subject-independent models.

User-Centric Design and Ethics

A core insight of the paper is that BCI is a User Experience (UX) problem, not just an engineering one. The authors emphasize "User-Centred Design," arguing that co-designing with end-users—such as medical staff for rehabilitation systems—is essential for long-term adoption.

They also address the "BCI Illiteracy" hurdle, noting that a significant portion of the population cannot use a BCI proficiently. To mitigate this, they suggest using pre-experiment questionnaires, like mental rotation tests, to predict performance and tailor training protocols. Furthermore, the paper delves into "Neurorights" and the impact of BCIs on personhood, raising critical questions about algorithmic responsibility and user agency.

Practical Implementation and Benchmarking

The paper provides a decision-making matrix for selecting components based on the system's goals:

  • Accuracy vs. Speed: If high information transfer rates are the priority, Reactive BCIs (like SSVEP) are recommended. If self-paced, natural control is needed, Active BCIs are superior.
  • Software Selection: BCI2000, OpenViBE, and BCILab are identified as the most popular platforms, with OpenViBE offering visual design tools for those who prefer not to write code.
  • Reproducibility: Researchers are encouraged to use established benchmarks like the MOABB (Mother of All BCI Benchmarks) to compare classifiers across diverse datasets.

Conclusion

While the authors acknowledge this chapter is a modest starting point, they point toward a future of "AutoBCIDesign." This envisioned tool would automatically select optimal sensors, algorithms, and paradigms based on specific user constraints. Although currently an ambitious goal, it represents the ultimate objective for making BCI implementation accessible to the wider scientific community.

Summary updated

BibTeX
@inproceedings{lotte2025how,
  author = {Fabien Lotte and Pauline Dreyer and Sébastien Rimbert and David Trocellier and Marc Welter},
  title = {{How to build a brain–computer interface from beginning to end}},
  booktitle = {Elsevier},
  year = {2025},
  doi = {10.1016/b978-0-443-24824-5.00004-1},
}
Figure