PurposeTo examine associations between fluctuating consciousness and Confusion Assessment Method for the Intensive Care Unit (CAM-ICU) assessments in stroke patients compared to non-neurological patients. Materials and methodsWe linked all recorded CAM-ICU assessments with corresponding Richmond Agitation Sedation Scale (RASS) measurements in patients with stroke or sepsis from a single-center ICU database. Fluctuating consciousness was defined by RASS variability using standard deviations (SD) over 24-h periods; regression analyses were performed to determine associations with RASS variability and CAM-ICU rating. ResultsWe identified 16,509 paired daily summaries of CAM-ICU and RASS measurements in 546 stroke patients and 1586 sepsis patients. Stroke patients had higher odds of positive (OR 4.2, 95% CI 3.3–5.5) and “unable to assess” (UTA; OR 5.2, 95% CI 4.0–6.8) CAM-ICU ratings compared to sepsis patients, and CAM-ICU-positive and UTA assessment-days had higher RASS variability than CAM-ICU-negative assessment-days, especially in stroke patients. Based on model-implied associations of RASS variability (OR 2.0 per semi-IQR-difference in RASS-SD, 95% CI 1.7–2.2) and stroke diagnosis (OR 2.7, 95% CI 2.0–3.7) with CAM-ICU-positive assessments, over one-third of probable delirium cases among stroke patients were potentially missed by the CAM-ICU. ConclusionsPost-stroke delirium may frequently go undetected by the CAM-ICU, even in the setting of fluctuating consciousness.