Interactive evidence · Synthetic data only

Data quality changes the model story.

A small, inspectable experiment showing why missingness, noise, imbalance and drift must be treated as part of model evaluation—not as cleanup after the result.

Demonstration boundaryNo uploads, patient data, proprietary data, diagnosis or medical recommendation. Every result comes from one deterministic synthetic operational signal.

01Inspect a clean baseline
02Introduce one quality problem
03Detect what changed
04Apply a declared remediation
05Compare coverage and evaluation

Choose a quality problem

Problem

Missing valuesDeterministic gaps remove blocks of samples.Spike noiseAdditive spikes corrupt individual samples.Class imbalanceEvent examples vanish from the calibration window.Distribution driftA level shift moves the held-out window.

SeverityModerate

Inspect the signal

Clean reference

Synthetic operational signal only. The chart is supplementary; equivalent values and comparisons appear in the tables below.

Baseline reading

No quality rule is triggered.

All 336 samples are present. The detector scores 0.911 accuracy on the untouched held-out window.

Compare the evidence

Coverage and model metrics are intentionally separated. When both states are available, the paired table scores them on the same held-out samples so cohort changes cannot masquerade as improvement.

Quality diagnostics

Diagnostic values for each visible pipeline state.
Measure	Clean
Completeness	100.0%
Outlier rate	0.0%
Level shift	-0.34
Calibration event share	17.9%
Held-out coverage	112 / 112

Detector evaluation

Full-state results. Coverage can differ, so use the paired table for like-for-like change.
Measure	Clean
Accuracy	0.911
Balanced accuracy	0.762
Precision	1.000
Recall	0.524
F1	0.688
Confusion counts	TP 11 · FP 0 · FN 10 · TN 91

Assumptions and provenance

What is generated

336 hourly samples across 14 days, with daily and weekly seasonality, nine fixed event excursions and seeded Gaussian noise with standard deviation 0.9. Seed 20260721 makes every run reproducible.

How it is evaluated

The first 224 hours fit a threshold detector. The remaining 112 hours are held out for evaluation against untouched synthetic labels. Results are descriptive, not evidence of clinical or operational validity.

Read the scenario definitions

Missing values

Corruption: Seeded block gaps (2–7 samples each) delete roughly 10%, 22% or 38% of the record depending on severity.
Detection: Completeness — defined samples divided by total samples; flagged below 99%.
Remediation: Bounded linear interpolation: gaps of at most five samples are filled from their surviving neighbours; longer gaps stay missing.
Limitation: Offline interpolation reads the sample after each gap — future information that a live system would not have. Interpolation recovers coverage, not information: paired before/after comparisons are made on the samples evaluable in both states.

Spike noise

Corruption: Seeded spikes hit roughly 3%, 6% or 10% of samples with magnitudes near 9–15 units, mostly upward.
Detection: Outlier rate — share of samples whose residual against the median of their four nearest neighbours exceeds 6 robust standard deviations (1.4826 × median absolute residual); flagged above 1%.
Remediation: A centred rolling median (window of five samples) replaces every value, suppressing isolated spikes.
Limitation: The smoother also blunts genuine event onsets, so recall can fall even as precision recovers.

Class imbalance

Corruption: A seeded draw discards 55%, 75% or 88% of event-labelled calibration samples; the held-out window is untouched.
Detection: Calibration event share — flagged when events fall below 12% of kept calibration samples (about 18% is expected).
Remediation: The discarded samples are not restored — they are gone. The remediation is a fitting-objective recalibration: the threshold is re-fitted on the same thinned data to maximise balanced accuracy instead of accuracy.
Limitation: This is a decision-rule change, not a data repair. Changing the objective trades overall accuracy for event recall — the comparison is shown honestly either way.

Distribution drift

Corruption: A constant shift of 2.4, 3.6 or 5.0 units is added to every held-out sample, as if the sensor were re-mounted mid-record.
Detection: Level shift — median of held-out observations minus median of calibration observations; flagged beyond ±0.75 units.
Remediation: Seasonal-naïve correction: each held-out observation is compared with the calibration observation exactly one weekly cycle (168 hours) earlier; the median of those paired differences is the estimated shift, subtracted from the held-out window.
Limitation: Offline method with disclosed leakage: estimating the shift reads the whole held-out window — future information a live system would lack. The calibration-only seasonal reference preserves the record's natural temporal structure, but the estimate carries residual noise and assumes one constant additive shift with stable weekly seasonality, so the result is an approximate correction, not unbiased recovery of the injected shift.

Read the metric and implementation notes

Accuracy is correct predictions divided by evaluated samples. Balanced accuracy averages event recall and non-event recall. Precision, recall and F1 are shown with confusion counts; a zero denominator is reported as undefined rather than zero.

Missing runs of at most 5 samples use bounded linear interpolation.
Spike remediation uses a centred rolling median of 5 samples.
Imbalance remediation changes the fitting objective; it does not recreate discarded samples.
Drift correction is offline and reads the held-out window, so it contains disclosed future-information leakage.

From demonstration to delivery

Evidence should make a system easier to question.

This lab is intentionally small. The same habits—explicit data boundaries, comparable evaluation, reproducible logic and visible limitations—shape the larger systems in the portfolio.

See selected systems Start a conversation