ai-algorithms-for-wave.md

🤖 AI / algorithms for better wave + prediction

by aoxilus 🥑 · what helps a 1-channel MindWave RF graph (Jul 2026).

Reality check

Goal	Realistic?
Cleaner raw graph / less noise	✅ Yes
Predict Attention / Meditation ~1–2 s ahead	✅ Weak but useful
Better blink vs noise	✅ Yes (best ROI)
Predict “thought color / L-R” from EEG	❌ No with 1ch

NeuroSky eSense is already a proprietary model (~1 Hz). AI on top helps smooth, forecast, classify blinks — it does not replace the chip.

---

Ranked options (best → heaviest)

1. 🥇 Blink: derivative peak/valley (classic DSP)

• Paper: configurable blink detector, single-channel BCI (FPGA, beats vendor software on latency)
• Idea: differentiate raw → find max then min on same slope (blink shape), not only amplitude vs baseline
• Fit for us: improve createBlinkFromRaw in JS — no training
• GitHub-ish: Matuteale/final-project — logistic regression on a moving window of raw for wink/not

2. 🥈 Graph: 1-D Kalman / EWMA + short forecast line

• Papers/repos: robust Kalman + ELM for attention · LowLatencyEEGFiltering (Kalman + TCN)
• Idea: filter display series; draw predicted next N samples as a dashed overlay
• Fit for us: ~30 lines of JS on wave.html (Kalman scalar or Holt linear). Runs at 512 Hz avg×5

3. 🥉 Short-horizon Attention forecast (AR / tiny MLP)

• Input: last 10–30 Attention values (~10–30 s)
• Output: Attention in +1 s / +2 s → early “CONCENTRADO / RELAJADO”
• Models that communities use: linear AR, logistic, small NN, SVM, XGBoost on features (RMS, band powers, Hjorth)
• LabVIEW/NeuroSky NN example: preprint — needs labeled sessions (we already export JSON)

4. Alpha / relax features (eyes closed vs open)

• Neural-Decoder — alpha 8–12 Hz + BiLSTM (OpenBCI, but same idea)
• Fit for us: we already FFT 0–50 Hz — add α power feature for RELAJADO (no heavy net first)

5. Deep denoise (later / offline)

• EEGdenoiseNet / Single-Channel-EEG-Denoise · EKFNet
• Strong papers, not first for browser realtime without a trained ONNX model

---

What we should try next (this repo)

Step	Algo	Where	Predicts
A	Kalman display + 200 ms forecast dash	src/wave.js	smoother graph
B	Blink slope detector (diff max→min)	study-lib / wave	fewer false blinks
C	α-band power → relax score	FFT already there	better RELAJADO
D	AR(5) on Attention series	wave state	“va a concentrarse”

Data we already have for training later: docs/studies/*.json, study/*.json (local).

---

Do NOT chase

1. LSTM to decode color/shape (null result in our exam).
2. Replacing eSense with a homebrew Attention without a big labeled set.
3. Heavy denoise nets before A–C above.

---

🥑 aoxilus