VoltEPIC · Explainable Prediction of Individual Concentrations

Four biomarkers, one scan — and the reason behind every call.

Ascorbic acid, serotonin, estradiol and melatonin, detected from a single square-wave voltammogram. Trained on 1,114 curves, checked on 125 it never saw, and explained with the methods that passed our own tests.

Open the analyzer How it works
  • Runs in your browser — nothing is uploaded
  • CSV or raw CHI760F files
  • ENG Peak Deconvolution · Team 2
Live readout loading…
0.0 Vpotential0.8 V
    Real held-out curves · shading: where the model looks
    1,239curves

    ACS-ML-Biomarker: 455 measurement blocks on P4VP- and PVDF-coated chips, four analytes, alone and in mixtures.

    125held out

    Curves from 46 blocks the model never saw in training. The library in the analyzer is exactly this set.

    —

    Held-out accuracy across the four analytes, recomputed in your browser from the stored predictions.

    0/ 500

    Decisions that differ between this page and the Python model it was exported from.

    How a decision is made

    From a current trace to four answers, in under a millisecond.

    Every step below is computed live from one held-out curve, with the same code the analyzer uses.

    1. 01

      Measure

      Square-wave voltammetry, 0.004–0.800 V in 4 mV steps: 200 currents.

    2. 02

      Denoise

      A wavelet threshold removes the noise and keeps the peak shape.

    3. 03

      Decompose

      Seven wavelet bands, 228 coefficients, from coarse shape to fine detail.

    4. 04

      Vote

      Four boosted forests, 300 trees each, build up the evidence tree by tree.

    5. 05

      Decide

      Each probability is compared with a threshold chosen on training data.

    Walk through the method
    Explainability

    It looks where the peaks are.

    Shuffle one 20 mV window across all held-out curves and measure how much worse the model gets: that is permutation importance, the most reliable of the four explanation methods we tested. On PVDF, the single most important window of each analyte sits inside that analyte's peak window.

      How we tested the explanations
      Permutation importance per window · PVDFthis model
      Each line is scaled to its own maximum. Bars under the axis: peak windows from the E3 test (blank-subtracted single-analyte means, PVDF).
      Built to be checked

      Claims you can verify on this site.

      Same answers as Python

      The wavelet transform and the 1,200 trees were reimplemented in JavaScript. On every page load the analyzer re-checks all 500 held-out decisions against the Python reference.

      Reference check →

      Explanations under test

      Four methods, five tests: does the map hold when you delete what it highlights, does it stay stable, does it sit on the peak, does it chase artefacts, does it vanish for a model that learned nothing?

      The five tests →

      Honest about limits

      One fitted model and one partition; replicates share a block; the model has only seen these chips and this instrument, and it cannot tell when a curve is from somewhere else.

      Model card →
      Analysis assistant · preview

      Ask the analysis in plain words.

      The analyzer answers questions about the curve in front of you — what was found, how sure the model is, which windows mattered — using only the numbers it computed. Today the answers are scripted; a language model will take over the same panel.

      Try the assistant
      AssistantDemo
      Which regions influenced the prediction?
      Ascorbic acid: 0.24–0.26 V, adding +3.0 log-odds, inside its peak window. Estradiol: 0.50–0.52 V (+2.9). Melatonin: 0.72–0.74 V (+2.7).
      How confident is the model?

      Bring a measurement.

      Drop a CSV or a raw CHI760F export into the analyzer, or start from one of the 125 held-out curves.

      Upload a file Random held-out curve