Summary

What we built

A complete SSVEP-BCI pipeline, end to end, on a four-file dataset:

  • Ch 1 — opened the four .mat files, decoded the eleven channels, and discovered that the trial-class mapping for CH11 lives in a screenshot rather than the description text.
  • Ch 2 — saw the SSVEP as a peak in raw PSD, with a working subject (subject 1 run 1) that lets the response stand visibly above the alpha rhythm.
  • Ch 3 — designed a 5–40 Hz bandpass + 50 Hz notch and showed honestly that on this file filtering is mostly cosmetic: it removes drift and would handle line noise if there were any, but doesn’t change the in-band spectrum.
  • Ch 4 — recovered the trial structure from CH10 transitions and built the canonical (20, 8, 1884) epoch tensor that every later chapter consumes.
  • Ch 5 — sampled the PSD at the four candidate frequencies (and their second harmonics) on each of the eight channels, then converted raw power into peak-vs-neighbourhood SNR features.
  • Ch 6 — built sine/cosine reference templates per candidate frequency and used CCA to score each epoch against them. FBCCA extended the idea across sub-bands.
  • Ch 7 — turned features into class labels, comparing CCA argmax (no training) against LDA on the spectral features and the CH11 baseline shipped with the data. CCA cleared 90 %; FBCCA hit 95 %; LDA-on-features and CH11 tied near 55 %.
  • Ch 8 — swept epoch length from 1 s to 7 s and showed that peak Information Transfer Rate sits at 2–3 s, well below where the accuracy curve plateaus.
  • Ch 9 — ran the pipeline on both subjects and observed that subject 1 (the one whose CH11 baseline was broken) actually has the cleaner SSVEP for our methods — confirming both subjects fit inside the distribution reported by Guger et al. (2012).

What was actually surprising

The book’s a-priori plan and what the data actually showed disagreed in three places worth re-stating:

  1. The shipped baseline lied about which subject was “easier”. Subject 2 run 2 was picked as the running example in Ch 1 because its CH11 LDA looked clean. Subject 1 run 1’s CH11 was stuck on class 3 — the “broken” file. With our own CCA classifier the ranking inverts: subject 1 saturates at 100 % by 4 s, subject 2 reaches that only at 7 s. Don’t let the shipped baseline pick your favourite subject.
  2. Filtering on this dataset is mostly cosmetic, not corrective. Butterworth filters are flat in the passband by construction, so the alpha rhythm sitting on top of 9–12 Hz stimulation frequencies survives filtering untouched. What handles alpha is features that care about phase (CCA), not preprocessing.
  3. For SSVEP, feature engineering is the classifier. A per-class CCA score is intrinsically class-comparable, so the “model” is just argmax. The work that distinguishes a 90 % pipeline from a 55 % one happens at the feature stage. Generic classifiers on top of generic features (LDA on amplitudes) leave most of the available structure on the table.

What we deliberately skipped

This was a single-dataset, single-pipeline walkthrough. The following are out of scope but real:

  • Subject-independent models — train on one user, test on another. Possible with template-matching methods that don’t need per-user calibration; needs many more subjects to evaluate.
  • Deep learning — EEGNet, FBCNN, transformer variants. Not obviously better than CCA on small SSVEP datasets, but the comparison would need careful regularization and a held-out test split.
  • Cross-session transfer — train on training_1, test on training_2 of the same subject. The two runs were recorded back-to-back, so this would mostly probe within-day stability; the more interesting question (across-day stability) needs different data.
  • Artifact rejection — ICA, regression-based EOG correction. SSVEP is robust enough that this rarely helps; for a paradigm like P300 the same shortcut wouldn’t work.
  • Online / asynchronous BCI — onset detection without a known trigger, idle-state rejection, dwell-time tuning. The offline accuracy and ITR numbers we reported assume an oracle stimulus onset.
  • Population-scale claims — anything beyond “we fit inside the published distribution” needs N ≫ 2 and pre-registered evaluation. Guger et al. (2012) with N = 53 is the right shape of study; a re-analysis of one dataset is not.

Where to go next

Concrete starting points that build on this book’s pipeline:

  • Switch to a larger benchmark. The Tsinghua BETA and Benchmark SSVEP datasets are open and large enough to test classifier choices, calibration regimes, and cross-subject transfer. The pipeline in this book translates directly — load, filter, epoch, score.
  • Try TRCA. Task-Related Component Analysis learns spatial filters per class from training data and tends to beat CCA / FBCCA when calibration data is available. A natural next chapter.
  • Calibrated vs calibration-free. Within the same dataset, compare CCA argmax (no calibration), TRCA (per-subject calibration), and a transfer-learning method that calibrates on others. The interesting question is whether the calibration cost is worth it for the user.
  • Measure end-to-end ITR. The Wolpaw formula uses T = epoch length here. In production, T includes onset detection, processing delay, network round-trip, and dwell time. A simple instrumented prototype would put real numbers on the gap between offline and online.

The pipeline you have at the end of Ch 9 is enough to do any of these — just point it at different data, a different classifier function, or a different evaluation loop.