Reproducible BCI research means publishing the pipeline, not describing it

A methods section is a summary written from memory weeks after the run. A study that carries its own stages, parameters and fold scores is the run itself.

The reproducibility problem in BCI research is not usually fraud, and it is rarely even carelessness. It is that the artifact we publish — a paper — is a lossy description of the thing that produced the result. The filter order, the rejection threshold, the fold assignment and the random seed do not fit in a methods paragraph, and by the time someone tries to reproduce the work, the script that knew them has been edited forty times.

BCILattice and BCINexus attack that from two ends. Locally, every stage of a study records what it ran with. On publication, the study carries those records with it, so a reader inspects the pipeline instead of reconstructing it.

Provenance is written as you go, not reconstructed after

Each analysis step reads a named stage, writes a named stage, and records its parameters. Nothing is destructive, so the input to any step is still there to re-run against. That turns "what did we do to this recording" from a question of memory into a property of the project file.

Because a step writes rather than overwrites, two settings can coexist. The comparison between a 0.5 Hz and a 1 Hz high-pass becomes a thing you can show rather than a thing you assert.

The guardrails catch the errors that survive peer review

The Results page carries a dedicated validation tab that reads the run design, infers the grouping, and states which comparisons the design supports. A result from a leaky split or an unnested search is still listed, with its real number and a flag — hiding it would just move the problem.

  • Grouped splits

    Keep a subject wholly in train or wholly in test. Trial-level random splits across a single subject are the single most common source of an unreproducible BCI accuracy.

  • Nested search

    An inner loop selects, an outer loop scores. The number that reaches a report came from folds no selection ever saw.

  • Leakage checks

    Fitting scalers, feature selectors or dimensionality reduction on the full dataset before splitting is flagged, not silently permitted.

  • Per-fold reporting

    A mean of 78% over folds of 95 and 61 is a different finding than one over folds of 79 and 77, and only one of them replicates.

A planned run often scores lower, and that is the point

Smart Training makes two claims and refuses a third. It claims provenance — every configuration decision derived from this cohort and this device, written down with its reason — and bounded effort, with the configurations tried and the rule that discarded them in the record.

It does not claim to make your model better. Its one-line contract is cited from every module that implements it: Smart Training may optimise, but never hide. If you want a figure that goes up, it is the wrong checkbox.

What a published study carries

Publishing on BCINexus sends the study record, and optionally its datasets and trained model weights. Reviewers work through a structured process — reviewers join, one claims the decision, comments accumulate, an outcome is recorded, quorum decides what happens next — rather than an email thread.

Published work can be corrected, disputed and, where warranted, retracted; retraction is the only exit from published, so the record of what was once claimed does not simply disappear. Studies can be cited, forked as the basis for new work, and linked to an existing paper by its DOI.

Reproducible does not mean public

Everything above works on a study that never leaves your machine. Provenance, guardrails and validation are local features; publication is a separate decision you make afterwards.

That matters for recordings under an ethics approval that does not permit sharing. A study can be fully reproducible inside a lab and never be published at all.

Questions

Does BCINexus issue a DOI for my study?

No. It links an existing DOI — resolved from Crossref or OpenAlex — to the study, so a published paper and the pipeline behind it point at each other. The DOI itself comes from your publisher or repository.

Can someone re-run my published study?

They can inspect the pipeline that produced it, including the stage parameters and fold-level results, and fork it as the starting point for their own work. Whether they can re-run it against the original data depends on whether you published the dataset with it.

What stops me publishing a result from a leaky split?

Nothing stops you, and that is deliberate — the validation tab flags it rather than blocking it, and the flag travels with the result. Suppressing the number would only relocate the problem to a place nobody can see.

Do I have to publish to benefit from this?

No. Provenance recording, the evaluation guardrails and the validation tab are local. Publication is a separate, explicit action.

Read the reference

Related

Run a study end to end, on your own machine

The free plan covers two studies, the full analysis suite, and 118 of the 119 ML blocks.