BIOSIGNAL / SURVEILLANCE STRESS-TEST LAB
A confident answer.
But can you trust it?
Put a signal classifier under pressure. Change the sample conditions, inspect its errors, and test when it should defer to a reviewer.
Simulated count data. Four invented channels represent a sequencing-style composition. This project tests statistical behaviour; it does not identify organisms, detect outbreaks or establish biological risk.
NEW / INDEPENDENT EVALUATION
Bring your own predictions.
Compare a baseline, inspect calibration and quantify review burden on batch-separated validation and test data.
Open evaluation workspace ↗Local CSV analysis · Research demo · Bundled example is synthetic
LIVE ANALYSIS
Validation results
Signal score distribution
Reference negative Reference positive
Errors versus deferral
Curve varies the margin; quality checks stay fixed. Low coverage means more review work.
SAME SAMPLES · DIFFERENT DECISIONS
Compare the trade-offs.
| Approach | Accuracy | Sensitivity | Specificity | Coverage | FP / FN |
|---|
The baseline thresholds channel A alone. The trained model uses three fractions and log read depth. “Model + defer” evaluates only retained predictions; withheld samples are neither correct nor negative.
AUDIT THE FAILURES
Every sample is inspectable.
| Sample / batch | Reference | Model score | Prediction | Decision |
|---|
TAKE THE EVIDENCE WITH YOU
Reproduce. Challenge. Improve.
Export the exact configuration, predictions and provenance behind this run.
METHODS / VERSION 1.0
A working benchmark, with inspectable limits.
What runs?
Python generates 1,920 labelled synthetic count samples across 240 collection batches. A standardised logistic regression is fitted on 144 batches. Validation and test each contain 48 different batches. Both model and channel-A baseline thresholds are selected on validation balanced accuracy.
The browser applies the exported Python coefficients to each sample. Stress transformations change observed counts while preserving the original latent reference label. No LLM or server is needed for inference.
When does it defer?
A prediction is withheld if depth is below 1,000, channel B is marked missing, known contamination exceeds 12%, any feature is more than four training standard deviations from its mean, or the score is within your selected threshold margin.
These are transparent heuristics, not a calibrated uncertainty guarantee. Unknown contamination can evade the QC flag. Confidence scores are not probabilities of a biological threat.
Reproduce the experiment
The research pack contains training code, pinned dependencies, data and tests. Apply for access below. Approved applicants can reproduce training with:
python source/train.py --out rebuilt node source/test.cjs python source/train.py --evaluate template.csv \ --model model.json
· · ·
What this cannot establish
There are no real sequencing reads, taxonomic assignments, pathogens or measured outbreaks here. Synthetic group separation prevents one form of leakage but does not prove external validity. Batch-bootstrap intervals in the Python benchmark describe the clean synthetic test set only.
Importing a CSV does not retrain the model. Its labels, feature mapping, quality metadata and applicability need independent review. No personal or patient data is required.
Training record & baseline benchmark
Loading…
Methods: scikit-learn logistic regression · training-set standardisation. Independent project by Prem Kumar Perumal.
