Association of admission acid-base phenotypes with delirium in cardiac critical illness: a multicenter cohort study

Acid-base disturbances frequently complicate cardiac critical illness. Although their associations with mortality have been described, their relationship with neurologic complications, particularly delirium, remains poorly defined. We analyzed data from 1259 adults with cardiac critical illness enrolled in an eight-center intensive care unit (ICU) registry (2011–2025). Using a physiological Boston approach, admission arterial blood gas profiles were categorized into six mutually exclusive phenotypes. Sequentially adjusted logistic regression models assessed associations with delirium, and separate adjusted models assessed associations with short-term mortality. Delirium was documented in 555 patients (44.1%). Compared with no acid-base disorder, admission respiratory alkalosis was associated with higher odds of delirium after multivariable adjustment (adjusted odds ratio [OR], 2.76; 95% confidence interval [CI], 1.67–4.54; p < 0.001). The association remained statistically significant after accounting for center-level clustering and ICU admission year (adjusted OR, 2.33; 95% CI, 1.20–4.52; p = 0.012). Each 5-mmHg decrease in PaCO₂ was associated with higher odds of documented delirium in Model 3 (adjusted OR, 1.07; 95% CI, 1.02–1.12; p = 0.004), but the association was not statistically significant after additionally accounting for center and admission year (adjusted OR, 1.06; 95% CI, 0.99–1.14; p = 0.078). No other phenotype remained significantly associated with delirium after full adjustment. Although metabolic acidosis had the highest crude ICU (22.6%) and in-hospital (29.8%) mortality, no acid-base phenotype remained significantly associated with mortality after adjustment for illness severity and mechanical circulatory support. Admission respiratory alkalosis was associated with delirium documented during the ICU stay in patients with cardiac critical illness. The continuous PaCO₂ association was less consistent across adjustment models. Residual confounding and uncertainty regarding the temporal sequence preclude causal inference.