Synthetic lab time series¶

openmed.training.synthetic.lab_timeseries generates offline numerical fixtures under caller-declared units, ranges, trends, variability, relative-day schedules, missing indices and additive intervention effects. No clinical reference ranges, physiological model, proprietary terminology or actual patient records are bundled. These are evaluator fixtures, not a claim of physiological fidelity.

Create LabSeriesSpec with a controlled analyte code, unit, count, lower/upper bounds and baseline. Optional LabIntervention(index, effect, intervention_code) changes the expected trajectory from its sample index onward. Every planned sample, including the full jitter envelope, must fit the declared range and step change limit. Impossible recipes reject; values are never silently clamped. Declared missing indices remain explicitly None while the underlying synthetic trajectory continues. Relative days can be shifted without changing intervals.

generate_lab_series(spec, seed=...) uses a private seeded RNG, records the complete recipe digest, and returns bounded immutable samples. It is deterministic within the documented Python runtime; it makes no cross-runtime random-number guarantee. validate_lab_series independently checks recipe binding, sample count/time, units, exact missingness, ranges, trends, intervention effects and change limits. Controlled error categories do not contain source values or private paths. Jitter and step checks compare representable floating-point interval endpoints, so small declared changes at large magnitudes are not rejected because of subtraction rounding. Numerical range checks remain inclusive and strict.

lab_negative_control(..., kind="range" | "unit" | "time" | "missingness") emits one deliberate corruption with negative_control split. The normal validator rejects that split and the underlying corruption independently. Keep those fixtures outside normal synthetic training splits to test the evaluator. Samples are limited to 10,000 and interventions to 1,000; no files or network requests are created implicitly. Recipe constraints are declarations rather than scientific evidence; clinical review is required for realistic parameter choices.