Synthetic IQC · AI/ML · Data Science · Training · Software Validation
IQC-Sim
Generate seed-reproducible longitudinal internal quality-control datasets with known ground-truth process states, analytical disturbances and Westgard-rule outcomes. IQC-Sim is designed for laboratory AI/ML and data-science research, algorithm benchmarking, anomaly-detection studies, QC software validation and training.
Known ground truthSeed reproducibilityLongitudinal QCUp to 100,000 rows
Synthetic data only. Not intended to establish laboratory QC limits, evaluate patient results, or substitute for laboratory-specific QC procedures.
Research and data-science applications
IQC-Sim creates labelled longitudinal QC datasets in which the generated process state and the independently derived rule-based QC status are both retained. This supports controlled method development where the underlying simulated truth is known.
AI/ML modellingSupervised classification, anomaly detection, time-series modelling and feature-ablation experiments.
Change detectionShift/drift recognition, change-point detection and sequential monitoring research.
Software validationPipeline regression tests, QC middleware checks and rule-engine benchmarking.
TrainingLaboratory education, competency exercises and reproducible teaching datasets.
Generated datasets model statistical QC behaviour. They should not be interpreted as validated representations of the error distribution of a specific analyser or laboratory.
1 · Simulation parameters
Therapeutic drugs are intentionally excluded. Rows without concrete numeric assignment parameters are also excluded. Each remaining source-table method is a separate option. The embedded values are numerical anchors for simulation, not laboratory-established QC limits.
2 · Run schedule
Allowed: 30–50,000 days
Allowed: 1–10/day, subject to the 100,000-row cap
Scheduled rows90 / 100,000
Max runs/day now10
Max days now50,000
Estimated CSV≈ 22 KB
Hard output limit: 100,000 rows. Days × runs/day must remain ≤100,000. At 50,000 days, a maximum of 2 runs/day is permitted; at 10 runs/day, a maximum of 10,000 days is permitted. File size is an estimate and varies with values and triggered-rule text.
Rule-clean creates a controlled background useful for labelled AI/ML challenge datasets by resampling ordinary scatter until selected multirules remain clear. Natural Gaussian retains unrestricted N(0,1) scatter, allowing incidental warnings/rejections for more naturalistic statistical experiments.
Modern evaluates rejection rules independently. Classic retains 1₂s as the warning gate for 4₁s and 10x.
3 · QC rule and event configuration
Select the rule outcomes evaluated in the generated dataset. Rejection rules may also be used to construct synthetic ground-truth disturbance episodes, allowing rule flags and QC status to serve as labels or features in downstream data-science workflows. 1₂s is warning-only and is never planted as an error episode or treated as a rejection. No quota is imposed; rules may overlap or arise incidentally. R₄s is only valid when both control levels are run.
Scope is intentionally restricted to the classic two-control Westgard multirule family: 1₂s, 1₃s, 2₂s, R₄s, 4₁s and 10x. 1₂s is a warning rule only; the other five are rejection rules. Three-control/N=3 variants (2of3₂s, 3₁s, 6x/9x configurations) and 7T are omitted. 8x/12x variants are also outside this deliberately narrow core implementation.
4 · Generate synthetic dataset
Filename: QCSim-Analyte-yymmddhhmmssfffff-seed. CSV uses UTF-8 with BOM and ASCII-safe SD-band/rule tokens for spreadsheet and data-science compatibility.
What is included in each dataset?
Run/date identifiers; known ground-truth process state; rule-based QC status; disturbance/event identifiers; target means and derived SD; generated L1/L2 values; deviations from mean and previous observations; standardized deviation indices; shift/drift magnitude; calibration/reset metadata; individual Westgard-rule flags; combined rule outcomes; warning/rejection indicators; and reproducibility metadata. The companion Markdown file records configuration, seed, event frequencies, rule overlaps and provenance. CSV categorical bands and combined rule labels use ASCII-safe tokens for reliable spreadsheet, Python/R and ML-pipeline import.
Generated dataset preview
Disturbance episodes count distinct planted process disturbances (targeted rule episodes + shifts + drifts). Targeted rule episodes count only episodes in which a rejection-rule signature was deliberately planted. Rejected runs are rule-engine outcomes and may occur during any disturbance.
Ground truth and rule-derived labels.GroundTruth_State records the process state deliberately generated by the simulator, whereas QC Status (WG Rules Based) is independently derived from the final rounded QC values as ACCEPT, WARNING or REJECT. Keeping them separate enables algorithms to be evaluated against known simulated ground truth rather than labels inferred from the same model input.
Technical validation: the frozen numerical engine correctly classified 49/49 predefined rule-engine challenges; 1,000 independently seeded simulations comprising 777,539 QC runs completed without a prespecified integrity failure; 18,043/18,043 targeted rejection signatures remained detectable after source-precision rounding; and identical settings/seeds produced byte-identical outputs through the 100,000-row test boundary. Version 1.0.1 changes export labelling/filenames only; the validated numerical simulation and rule engine are unchanged.