Dynamic Risk Stratification for Short-Term Deterioration After Acute Stroke: Incremental Value of Radiology-Report-Derived Neuroimaging Phenotypes Beyond Dynamic ICU States, a retrospective cohort study with temporal validation

Background. Prediction models after stroke increasingly add neuroimaging information to bedside data. Whether findings documented in radiology reports improve short term risk stratification beyond a patient's evolving physiological state is unknown. Methods. We studied 4,827 adults with acute ischaemic stroke, intracerebral haemorrhage or subarachnoid haemorrhage in MIMIC IV (2008 to 2019). At eight landmarks from 6 to 48 hours after ICU admission, models ranked patients by their risk of deterioration or death within 24 hours. The baseline model combined demographic and clinical characteristics, a previously validated dynamic ICU state representation, and the physiology, treatment and imaging status at each landmark. A locally run language model extracted imaging phenotypes from 6,992 head CT and MRI reports; three phenotypes and four location categories met a pre specified reliability criterion for automated extraction in independent validation and entered the model. The primary outcome was the difference in the proportion of events captured among the 10% of patients ranked highest at each landmark, in temporal validation (2017 to 2019; 8,923 patient landmarks, 666 events). Results. Among the highest ranked 10%, adding imaging phenotypes captured no additional event (348 of 666 with and without them): a difference of 0.00 percentage points (95% CI -2.20 to +2.01). Discrimination (AUROC 0.855 vs 0.849), calibration and net benefit were similar. No prespecified sensitivity analysis, stroke subtype or availability-adjusted analysis showed statistically supported incremental capture. In a post hoc redundancy analysis, imaging phenotypes improved capture when added to demographic and clinical characteristics alone (+12.91 points, 7.54 to 17.81), but this gain was no longer evident once the dynamic state representation was included. Conclusions. Neuroimaging phenotypes derived from routine radiology reports were associated with short-term deterioration, but we found no evidence that they added risk stratification value beyond the dynamic ICU state and contemporaneous physiology.