GERO-TWIN synthetic intensive care glycaemia cohort with latent ground truth for evaluating clinical digital twins

A synthetic cohort of 300 intensive care stays in adults aged 65 and over receiving insulin, generated by a mechanistic glucose and insulin model with per patient parameter jitter and a saturable clearance term. The distinguishing feature is that the latent generating states are retained. File 05 holds the true insulin sensitivity and the noise free glucose trajectory behind every observation, quantities that cannot exist in a patient record. Most synthetic clinical datasets are built to stand in for patient data and are judged by how closely they resemble it. This one is built for the opposite reason, to supply a known answer against which an analysis can be checked rather than compared with another estimate. Hypoglycaemia in the cohort is driven principally by drift in insulin sensitivity, which varies about fourfold within a patient, and that is the parameter a filter has to track. The deposit contains the stay table, the glucose measurement record, insulin and nutrition records, two forecast tables from a digital twin run with and without data assimilation, a data dictionary and the generator. It also contains a paired protocol counterfactual in which all 300 patients were re-run under three insulin protocols with every other random draw held fixed, included because it documents a boundary of the dataset. The cohort supports questions about estimation and evaluation and does not support questions about dosing protocol, which the counterfactual demonstrates rather than assumes. Truth was generated with a saturable clearance term the model under test does not contain, so the misspecification is deliberate. No patient data were used at any stage, there is no source cohort and no re-identification surface, and the whole cohort regenerates deterministically from seed 20260803.