Filter Sandbox

What it does

Drag cutoffs, order and transition bandwidth on a real 10-second EEG trace and watch four views update live: the filtered trace over the original, the spectrum before and after, the filter’s impulse response and its step response. Switch between FIR and IIR designs and between causal and zero-phase application. A canned electrode pop shows ringing; a blink shows what a high-pass does to a slow transient; a line-noise trace with harmonics shows what a notch does and does not remove. All filtering runs in the browser using the site’s own DSP library (src/lib/dsp/), which is tested against SciPy reference outputs.

Controls

Control	What it sets
Trace	Which recording to filter: clean eyes-closed (occipital), electrode pop, blink, line noise with harmonics, or an 8-channel segment
Filter type	Low-pass, high-pass, band-pass, band-stop (notch in notch mode)
Cutoff(s)	Low and/or high cutoff in Hz
Order / transition band	IIR order, or FIR transition bandwidth (which sets filter length)
FIR / IIR	Windowed-sinc FIR (MNE-style firwin) or Butterworth IIR (second-order sections)
Causal / zero-phase	Single forward pass, or forward–backward (filtfilt) application
Notch width, harmonics	(notch mode) width of the notch and how many harmonics to remove
Pipeline stage	(pipeline mode) where the filter sits relative to epoching and the ICA fit

What to look for

basic

• A sharper cutoff (higher order or narrower transition band) produces a longer, more oscillatory impulse response.
• Zero-phase filtering spreads the ringing around the electrode pop to both sides; causal filtering puts it only after, and delays the whole trace.
• Raising a high-pass from 0.1 Hz toward 1 Hz and above visibly reshapes the blink and adds opposite-polarity deflections around it.
• The spectrum shows what the filter removed; the trace shows what it did to the timing.

notch

• A narrow notch removes the 60 Hz line and leaves the flanks intact; a wide or high-order notch carves a hole and rings around the pop.
• Harmonics below the Nyquist frequency need their own notches; those above it cannot be addressed at their true frequency.

pipeline

• The same high-pass applied to a short epoch shows edge artifacts that the continuously filtered version does not.
• The 1 Hz “ICA-fit” copy and the 0.1 Hz “analysis” copy are different data; the unmixing matrix travels between them, the filtered signal does not.

causal

• Causal group delay is visible as a lag between original and filtered traces; it grows with FIR length and with IIR order at low frequencies.

Used in

• L1.5 Filters: FIR, IIR, and what they do to your data (basic)
• L1.6 Line noise (notch)
• L2.4 Filtering in practice (pipeline)
• L7.3 Real-time processing and neurofeedback (causal)

Data provenance

All traces are derivatives (cropped, single-channel or channel-subset, unfiltered unless stated) of public datasets; the exact subject, run, start time and license of each trace are read from the asset sidecar and shown in the frame’s provenance line.

• Clean eyes-closed (occipital), electrode pop, blink, and the 8-channel segment: ds-eegbci (EEG Motor Movement/Imagery Dataset; 160 Hz; recorded with no hardware filters; ODC-By 1.0 on PhysioNet, CC0 on the OpenNeuro mirror). Subject and run: TODO(confirm) from data/manifest.json.
• Line noise with harmonics: ds-iowapd (Iowa Parkinson’s resting EEG; 500 Hz; 60 Hz mains; CC0). Subject: TODO(confirm) from data/manifest.json.
• If no real electrode pop could be found, the pop trace is a labelled synthetic derivative (synthetic: true in its sidecar, with derived_from and modifications), and the frame says so.

Every asset is listed in data/manifest.json with its SHA-256 and generating script.

Open the code

site/src/components/widgets/w-filter-sandbox/ — Widget.svelte, index.ts, README.md. Repository link: TODO(confirm) (GitHub org/repo, §13 item 3).