NeuroBack tracks the brain's alpha rhythm in real time and delivers a light flash at a chosen phase of every cycle. It is an open-source, closed-loop EEG system for the OpenBCI Cyton. Research groups (PLOS Biology 2024) and consumer sleep headbands (Elemind) use the same technique, phase-locked sensory stimulation.
Open the live demo → It runs in the browser on recorded EEG, with flash times computed by the real engine. No installation needed.
It started as a 2023 university project at ITBA (Buenos Aires) that reproduced Huang et al. (2019). In 2026 it was rebuilt from the ground up:
- Modern phase tracking: the endpoint-corrected Hilbert transform.
- A reproducible benchmark: real EEG from 109 people.
- Proper controls: sham-style control conditions in the protocol.
- Safety and privacy built in.
- An honest post-mortem of what the original got wrong.
- Hits the target phase on real EEG.
- Median phase-locking value 0.74 across 109 people, eyes closed. The 2023 rule scored 0.42 with its input bug fixed and 0.08 as shipped.
- ecHT beat the 2023 rule in 109 of 109 people.
- The four target phases 0°, 90°, 180° and 270° land within 3° on average.
- Locks hardest when alpha is strong. Phase-locking reaches 0.94 in each person's strongest-alpha quartile, and an amplitude gate skips cycles without a clear rhythm.
- Robust to real hardware timing.
- Every sample is used: the 2023 loop silently dropped all but the last sample of each USB burst.
- Each flash time is refined as new data arrive.
- A measured output latency is compensated, and 3 ms of timing jitter costs only 0.01 in phase-locking value. For scale, the Cyton's measured jitter with the FTDI fix is 2.1 ms (Knierim et al., 2023).
- Controls built in. The phase sweep repeats four target phases in randomised order, alongside open-loop flicker at the individual alpha frequency (IAF) and random-phase flashes, with alpha measured in the stimulus-free gap after each block. A permutation test reports whether alpha really depends on phase.
- Safe and private by default.
- Stimulation stays locked until photosensitivity screening and consent are recorded, and there is an Esc emergency stop.
- Participants exist only as random pseudonyms, and every SQL query is parameterised.
- "Delete participant" removes everything.
- Runs without hardware. Simulated alpha, playback of public EEG, BrainFlow's synthetic board, or the online demo; 38 tests and CI.
- Modern interface, Python timing. A browser interface served locally by a FastAPI backend. The closed loop and the trigger scheduler run in Python, so the browser never sets the timing.
git clone https://github.com/abrilcode/neuroback && cd neuroback
pip install -e ".[all]"
neuroback demo # headless closed loop on simulated alpha, prints accuracy
neuroback app # interface in the browser, simulated alpha (on-screen LED)
neuroback app --board playback # replay real EEG (PhysioNet subject 1, downloaded on demand)
neuroback benchmark # reproduce the 109-person benchmark (~140 MB download, ~2 min)With hardware (OpenBCI Cyton + Arduino with firmware/neuroback_trigger):
neuroback app --board cyton --port /dev/cu.usbserial-XXXX --channel 8 \
--trigger-port /dev/cu.usbmodemXXXX --latency-ms 12Read docs/hardware.md first: it covers the electrode montage, the FTDI latency fix, LED safety, and how to measure --latency-ms with a photodiode.
The loop runs once per alpha cycle:
- Recording: an electrode at Oz, over the visual cortex, picks up the alpha rhythm. The OpenBCI Cyton digitises it at 250 Hz and radios it to the laptop.
- Timing: NeuroBack estimates the alpha phase and tells an Arduino exactly when to pulse a diffused LED.
- Closing the loop: the flash reaches the retina and travels through the thalamus (LGN) to the visual cortex (V1), where the next alpha cycle is measured.
Parts, wiring and latency calibration are in docs/hardware.md.
flowchart LR
A[Scalp EEG at Oz<br/>OpenBCI Cyton, 250 Hz] -->|BrainFlow, every sample| B[Ring buffer<br/>0.5 s]
B --> C[ecHT phase estimate<br/>at the newest sample]
C --> D[Project to next target phase<br/>at the IAF, minus output latency]
D -->|refined every chunk| E[Trigger scheduler<br/>sub-ms spin-wait]
E -->|'F' over serial| F[Arduino pulse<br/>LED 20 ms]
F -->|light| G((Participant))
G --> A
- Calibrate. Record 60 s eyes open and 60 s eyes closed. NeuroBack finds the individual alpha frequency after removing the 1/f background, to about 0.05 Hz; the 2023 grid was 1.34 Hz. It also checks that alpha rises when the eyes close.
- Phase sweep. Four target phases plus two controls with the same flash rate, each repeated in randomised blocks (20 s on, 10 s off). A cosine fit with a permutation test checks whether post-block alpha depends on phase.
- Stimulate. Pick a target phase. The app shows where each flash actually landed one second later, with live phase-locking, mean error and flash rate.
 |
 |
 |
| Calibrate: IAF after 1/f removal, eyes-closed vs eyes-open spectra | Phase sweep: randomised targets and flash-rate-matched controls | Results: post-block alpha by condition and a permutation test |
The interface is a local web app: neuroback app starts a FastAPI server and opens the browser. Python runs acquisition, phase estimation and the sub-millisecond trigger scheduler; the browser only draws state frames (20 per second) and sends commands over a WebSocket. The online demo uses the same interface on a replay of PhysioNet subject S004. Its flash times and sweep were precomputed by the same engine (scripts/build_demo.py).
Each phase-tracking method processed the 1-minute eyes-closed recordings of all 109 people in the PhysioNet EEG Motor Movement/Imagery Database as a live stream:
- channel Oz, resampled to the Cyton's 250 Hz, delivered in 4-sample (16 ms) chunks. That is a middle ground: with the 1 ms FTDI fix the Cyton delivers samples almost one at a time, while its default buffering sends ~60-sample bursts that would make the 2023 dropped-sample bug far worse;
- aimed at the alpha peak;
- every flash scored afterwards against the non-causal Hilbert phase.
Settings were tuned on synthetic data only, so the real EEG is held out. Full method: docs/benchmark.md.
| Method | Phase-locking value, median (IQR) | Spread (circular SD) | Bias | Cycles stimulated | With 3 ms jitter |
|---|---|---|---|---|---|
| 2023 code, as shipped (oracle delay) | 0.08 (0.04–0.11) | ±129° | 11.4° | 25% | 0.07 |
| 2023 rule, input bug fixed (oracle delay) | 0.42 (0.31–0.53) | ±75° | 1.0° | 100% | 0.41 |
| AR forecast (phastimate-style) | 0.64 (0.52–0.78) | ±54° | 4.0° | 83% | 0.63 |
| ecHT (NeuroBack 2026) | 0.74 (0.67–0.83) | ±44° | 4.2° | 82% | 0.73 |
The 2023 rows get an oracle delay, the best possible one for each person. That is more generous than the original calibration sweep, which could not find it (see the post-mortem).
Version 1 was a real end-to-end system: Cyton acquisition, real-time filtering, an Arduino-driven LED, a PyQt6 GUI and a session database. Re-reading it in 2026 turned up bugs that meant the core result was never actually measured. The full audit, reproduced from the original files, is in docs/postmortem-2023.md.
| 2023 | 2026 |
|---|---|
| Fed only the last sample of each USB burst to the filter, giving near-random phase (0.08) | Every sample goes into the estimator (ecHT 0.74) |
| Rising zero-crossing on a causal ±1 Hz band-pass: ±90° error at 1 Hz off-centre, ~225 ms group delay | Endpoint-corrected Hilbert transform, with optional AR forecasting |
Final stage passed the index of the best delay (0–10) to time.sleep as seconds |
One scheduler thread; deadlines in seconds by construction, refined every chunk |
np.delete result discarded, so files repeated samples 2.5–4.5×, and per-delay windows read the wrong blocks |
Each sample stored once (.npz + JSON events) |
| 11 delays × 20 s, once each, in ascending order, so drift looked like a phase effect | Randomised repeated blocks, two flash-rate-matched controls, permutation test |
| IAF on a 1.34 Hz grid (only 8.02, 9.36, 10.70 or 12.03 Hz possible) | 1/f-corrected peak on a ~0.05 Hz grid, plus an eyes-closed reactivity check |
| Patient names as SQL table names via f-strings, names and EEG stored together | Random pseudonyms, fixed schema, parameterised queries, erasure |
| No photosensitivity safeguard for a ~10 Hz flashing LED | Screening + consent gate, warning, Esc emergency stop |
Phase-locked sensory stimulation is a real and active field. Sounds timed to the alpha phase shift alpha power and frequency depending on phase (Hebron et al., PLOS Biology 2024). A headband that plays sounds out of phase with alpha shortened sleep onset by 10.5 min versus a no-sound week (Bressler et al., Sci Rep 2024). NeuroBack's visual method comes from Huang et al. (NeuroImage 2019), which so far has been replicated only by its own group.
Clinical claims are a different matter. The 2023 slides suggested benefits for ADHD, anxiety, Parkinson's disease and migraine. Recent evidence does not support those claims for alpha protocols:
- ADHD: a 38-trial meta-analysis found SMD 0.04 when raters were probably blinded (Westwood et al., JAMA Psychiatry 2025).
- Children, alpha training: a 2026 double-blind trial of personalised upper-alpha neurofeedback found no difference from sham (Wang et al., J Child Psychol Psychiatry 2026).
- Alpha rising is not evidence of control: a preregistered study found alpha rose during a session whether the feedback was real or sham (Maaz et al., Communications Psychology 2026).
So NeuroBack is presented for what it is: a research and teaching instrument for closed-loop, phase-locked stimulation, with measured engineering performance. It is not a treatment. The full review, including where neurofeedback does show specific effects (closed-loop motor rehabilitation after stroke, small effects for standard ADHD protocols), is in docs/science.md.
Warning
NeuroBack flashes light at 8–13 Hz. Flicker in this range can trigger seizures in people with photosensitive epilepsy, sometimes with no prior history. Screen every participant, use a diffused LED at modest brightness, keep the Esc stop within reach, and stop at the first sign of discomfort. Research and teaching use only. Not a medical device, and no diagnostic or therapeutic claims are made. See docs/safety-and-privacy.md.
src/neuroback/
phase.py ecHT, AR forecast, and the 2023 rule (for comparison)
closedloop.py engine: phase -> next target time, latency compensation, controls
dsp.py IAF (1/f-corrected), band power, offline ground-truth phase
sources.py Cyton / BrainFlow, simulated alpha, playback
stimulators.py serial trigger + sub-millisecond scheduler
protocol.py randomised phase sweep with controls
analysis.py circular statistics, cosine fit, permutation test
storage.py pseudonymous SQLite store with erasure
safety.py photosensitivity screening
session.py session state machine (screening gate, calibration, sweep, stimulation)
web/ FastAPI server + browser interface (HTML, CSS, canvas)
bench/ PhysioNet loader + phase benchmark
firmware/ Arduino trigger (LED + TTL marker for the Cyton)
scripts/ figure + illustration generation (every image is reproducible)
docs/ benchmark, science review, hardware, protocol, safety, 2023 post-mortem
docs/demo/ online demo (GitHub Pages), built by scripts/build_demo.py
legacy/2023/ the original 2023 code, unchanged
- Hardware validation: photodiode loop-back measurements of latency and jitter on a Cyton, published in
docs/. - Auditory pulses as an alternative stimulus: this is the version independently replicated, and it avoids photosensitivity risk.
- State-space phase estimator (Wodeyar et al., 2021) with credible intervals.
- Lab Streaming Layer output and XDF recording for MNE-LSL / LabRecorder workflows.
- A preregistered, sham-controlled pilot (CRED-nf checklist) before any behavioural claim.
Work that uses NeuroBack should cite the software (CITATION.cff) and the methods it builds on: Huang et al. 2019; Schreglmann et al. 2021; Zrenner et al. 2020. Cite the dataset too: Schalk et al. 2004; Goldberger et al. 2000.
Created by Abril Herrlein. NeuroBack began in 2023 as a course project for Instrumentación Biomédica II at the Instituto Tecnológico de Buenos Aires (ITBA) and was rebuilt as v2 in 2026.
- Huang G. et al. (2019). A novel training-free externally-regulated neurofeedback (ER-NF) system using phase-guided visual stimulation for alpha modulation. NeuroImage 189, 688–699. doi:10.1016/j.neuroimage.2019.01.072
- Schreglmann S.R. et al. (2021). Non-invasive suppression of essential tremor via phase-locked disruption of its temporal coherence. Nature Communications 12, 363. doi:10.1038/s41467-020-20581-7
- Zrenner C. et al. (2020). The shaky ground truth of real-time phase estimation. NeuroImage 214, 116761. doi:10.1016/j.neuroimage.2020.116761
- Hebron H. et al. (2024). A closed-loop auditory stimulation approach selectively modulates alpha oscillations and sleep onset dynamics in humans. PLOS Biology 22, e3002651. doi:10.1371/journal.pbio.3002651
- Bressler S. et al. (2024). A randomized controlled trial of alpha phase-locked auditory stimulation to treat symptoms of sleep onset insomnia. Scientific Reports 14, 13039. doi:10.1038/s41598-024-63385-1
- Westwood S.J. et al. (2025). Neurofeedback for attention-deficit/hyperactivity disorder: a systematic review and meta-analysis. JAMA Psychiatry 82(2), 118–129. doi:10.1001/jamapsychiatry.2024.3702
- Maaz J. et al. (2026). Alpha power increases spontaneously during a neurofeedback session. Communications Psychology 4, 75. doi:10.1038/s44271-026-00431-w
- Knierim M.T., Bleichner M.G., Reali P. (2023). A systematic comparison of high-end and low-cost EEG amplifiers for concealed, around-the-ear EEG recordings. Sensors 23, 4559. doi:10.3390/s23094559
- Schalk G. et al. (2004). BCI2000: a general-purpose brain-computer interface (BCI) system. IEEE Transactions on Biomedical Engineering 51, 1034–1043. doi:10.1109/TBME.2004.827072
- Goldberger A.L. et al. (2000). PhysioBank, PhysioToolkit, and PhysioNet. Circulation 101, e215–e220. doi:10.1161/01.CIR.101.23.e215
Full reference list: docs/science.md.