Background and Objectives: Delayed cerebral ischemia (DCI) is a major determinant of poor outcome after aneurysmal subarachnoid hemorrhage (aSAH), yet early risk prediction remains difficult, particularly in sedated or ventilated intensive care patients. We evaluated whether a large language model (LLM) could predict DCI from phase-based dynamic clinical data, compared with intensive care specialists. Materials and Methods: In this single-center, retrospective study, 216 consecutive patients with aSAH were assessed at three predefined phases of accumulating clinical data (day 1; days 1 + 3; days 1 + 3 + 5). For each patient–phase, an LLM (ChatGPT, GPT-5.5 Thinking) and two blinded intensive care specialists predicted DCI risk using only the data available up to that time point. DCI was adjudicated by a blinded three-member panel. Discrimination was assessed by the area under the receiver operating characteristic curve (AUC), with non-inferiority defined a priori as Δ = 0.10. Calibration, decision-curve analysis, and reproducibility were also assessed. Results: Of 216 patients, 60 (27.8%) were DCI-positive and 15 were indeterminate; the primary sample comprised 201 patients. In the prespecified primary analysis, the LLM met the non-inferiority criterion relative to both specialists across all three phases (LLM AUC 0.703–0.747; specialists 0.712–0.764); however, in an equal-granularity sensitivity analysis, non-inferiority remained supported only in Phases 2 and 3 and was not demonstrated in Phase 1. Discrimination increased numerically as data accumulated (Phase 1 vs. 3, p = 0.072), an increase that was attenuated in a landmark-restricted analysis accounting for DCI-onset timing. The LLM showed the lowest false-reassurance rate (3.3–11.7%), reflecting a more cautious threshold rather than better discrimination. Confidence did not reliably indicate accuracy (~25% of high-confidence predictions were wrong); outputs were highly reproducible (Fleiss κ 0.887; intraclass correlation coefficient (ICC) 0.968). Conclusions: An LLM achieved DCI discrimination that was non-inferior to—but not better than—that of experienced specialists; its unreliable confidence scores support clinician-supervised rather than autonomous use.

Background: Weaning from mechanical ventilation remains a complex and failure-prone process, with extubation failure rates reaching up to 30%. Conventional indices inadequately capture the multifactorial physiology underlying weaning failure. This study aimed to evaluate whether a multimodal ultrasound approach could improve the identification of mechanisms and prediction of extubation outcomes. Methods: In this prospective pilot observational study, adult mechanically ventilated patients with preserved left ventricular ejection fraction (LVEF ≥ 50%) undergoing spontaneous breathing trials (SBT) were included. Multimodal ultrasound assessment—including transthoracic echocardiography (TTE), lung ultrasound (LUS), diaphragmatic ultrasound (DUS), and venous excess ultrasound (VExUS)—was performed at two predefined time points. Conventional respiratory mechanics parameters were recorded concurrently. The primary outcome was a composite of SBT failure (permanent weaning failure) or reintubation within 48 h. Results: A total of 27 patients were included, of whom 8 (29.6%) experienced extubation failure (5 permanent SBT failure, 3 post-extubation reintubation). Respiratory system compliance showed consistent associations with extubation failure across both ROC and regression analyses (AUC 0.806, 95% CI 0.611–1.000; cutoff ≤ 45 mL/cmH₂O; sensitivity 88%; specificity 74%; NPV 93%). Diaphragm excursion was significantly lower in the failure group (p = 0.042) and showed useful predictive performance (AUC 0.750, 95% CI 0.565–0.935; cutoff ≤ 24 mm; sensitivity 100%; specificity 58%; NPV 100%). Lung ultrasound, VExUS, and echocardiographic parameters did not demonstrate significant predictive value. Given the limited number of outcome events (n = 8) and events-per-variable ratio of 4.0 (EPV = 4.0), all multivariable findings are hypothesis-generating. Conclusions: In this prospective pilot study, respiratory system compliance and diaphragm excursion were associated with extubation failure in patients with preserved left ventricular function, while echocardiographic indices, LUS, and VExUS grading did not demonstrate significant predictive value. These hypothesis-generating findings suggest that impaired diaphragmatic function and reduced compliance may be more closely associated with weaning failure than cardiopulmonary congestion parameters. However, given the small sample size, low EPV, and single-centre design, all findings require validation in larger multicentre studies including patients with impaired systolic function.

Background/Objectives: Postoperative intensive care unit (ICU) disposition after anatomical lung resection is usually planned preoperatively, but the final care pathway may be substantially influenced by intraoperative events. We evaluated actual postoperative ICU admission among patients with a documented preoperative ICU monitoring recommendation and compared preoperative-only and perioperative triage approaches. Methods: In this retrospective single-centre cohort study, 1060 adults undergoing elective anatomical lung resection for non-small cell lung cancer (NSCLC) between January 2019 and December 2025 were screened; 159 patients with a documented preoperative ICU monitoring recommendation constituted the analytical cohort. A clinically pre-specified primary perioperative model incorporating operative duration, intraoperative complication, chronic obstructive pulmonary disease (COPD), and pre-existing arrhythmia was compared with a preoperative-only model and with an exploratory perioperative ICU triage score. Results: Actual postoperative ICU admission occurred in 45 patients (28.3%). Operative duration (adjusted odds ratio [OR] 1.012 per minute; 95% confidence interval [CI], 1.005–1.018; p < 0.001) and intraoperative complication (adjusted OR 15.002; 95% CI, 3.738–60.210; p < 0.001) were significantly associated with actual postoperative ICU admission. The primary perioperative model achieved an AUC of 0.802 (95% CI, 0.717–0.876), compared with 0.759 for the exploratory perioperative triage score and 0.665 for the preoperative-only model. Conclusions: Fewer than one-third of patients with a documented preoperative ICU monitoring recommendation underwent actual postoperative ICU admission. In this selected cohort, perioperative reassessment incorporating intraoperative information showed higher apparent discriminative performance than the preoperative-only approach while the exploratory score showed intermediate, hypothesis-generating performance. Because the outcome reflected observed institutional ICU disposition rather than independently adjudicated ICU-level care requirement, prospective multicentre validation using predefined ICU admission criteria is required before clinical implementation.

Background/Objectives: In aneurysmal subarachnoid hemorrhage (aSAH), intensive care unit (ICU) deaths may follow neurological or systemic pathways, but the relationship between early prognostic measures and the mechanism of death is unclear. Methods: We retrospectively reviewed 261 consecutive adults with aSAH admitted to the ICU between January 2020 and June 2026. World Federation of Neurosurgical Societies (WFNS) grade, modified Fisher scale and Subarachnoid Hemorrhage Early Brain Edema Score (SEBES) were assessed at admission; the neutrophil-to-lymphocyte ratio (NLR) and C-reactive protein (CRP)-to-albumin ratio (CAR) came from the first routine blood sample; and Acute Physiology and Chronic Health Evaluation II (APACHE II) and Sequential Organ Failure Assessment (SOFA) came from the first 24 h. Brain death was assigned to the neurological category by a predefined rule; the remaining deaths were adjudicated independently by two intensivists. Results: Eighty-six patients (33.0%) died: 45 (52.3%) neurological, 37 (43.0%) systemic/medical and 4 (4.7%) mixed/undetermined; 94.6% of systemic/medical deaths were sepsis-related. Agreement among the 54 independently adjudicated deaths was high (κ = 0.829). CAR discriminated toward systemic/medical death (AUC 0.671; 95% CI 0.552–0.791), whereas APACHE II (0.702) and SOFA (0.682) discriminated toward neurological death; WFNS (0.643) and SEBES (0.638) showed the same direction. NLR did not discriminate mechanism of death in this cohort (0.449). CAR was very highly correlated with CRP (ρ = 0.994). Conclusions: Among aSAH decedents, early prognostic measures were associated with the mechanism of death in opposite directions; the systemic/medical association predominantly reflected sepsis-related deaths. Because the analysis was restricted to decedents and time to death differed by mechanism (7 versus 26 days), the findings are exploratory and do not establish cause-specific risk prediction in the overall aSAH population.

Background and Objectives: Apnea test criteria differ between countries and have changed within them. The 2010 American Academy of Neurology guideline accepted a PaCO₂ ≥60 mmHg or a rise of ≥20 mmHg above baseline; the 2023 AAN/AAP/CNS/SCCM consensus guideline requires both together with an arterial pH <7.30; and the Turkish national standard, restated in December 2025, requires both carbon dioxide conditions but no pH target. We examined how the same real-world apnea tests are classified under these three criteria, and what accounts for any disagreement between them. Materials and Methods: All three criteria were applied to every apnea test performed during brain death determination in a tertiary intensive care service between 2019 and 2026 (n = 331). The primary endpoint was the 2023 consensus criterion. Determinants were modelled with Firth penalised logistic regression, with the Turkish criterion as a negative control. Results: 237 adults were evaluable under all three criteria. The proportion of tests meeting the criterion fell from 98.3% (AAN 2010) to 87.8% (Türkiye 2025) to 78.1% (2023 consensus); no test gained positivity at any step, and 48/237 (20.3%) were positive under at least one criterion but not all three. All 23 tests that met the Turkish criterion but failed the 2023 criterion had begun with a pre-test pH above 7.45. Among the 118 tests begun within the pH window that guidelines require, the two criteria were met by the same patients (106/118 (89.8) each). Pre-test bicarbonate was associated with meeting the 2023 criterion (OR 0.84 per mmol/L, 95% CI 0.78–0.92) but not with the Turkish criterion (OR 1.09, 95% CI 0.98–1.20). Conclusions: In this cohort, disagreement between contemporary apnea test criteria was determined mainly by whether the pre-test acid–base requirement had been met, rather than by the criteria themselves.

Background: First-pass failure during ultrasound-guided internal jugular vein (IJV) catheterisation is common in critically ill adults and is associated with mechanical complications, yet no model has been developed to estimate its probability before the procedure. We therefore set out to develop and internally validate a preliminary prediction model, ROTATE-ICU. Methods: We conducted a prospective, single-centre, observational cohort study and included 120 of 156 assessed procedures, taking one episode per patient. The pre-specified model combined IJV anterior–posterior diameter, the position of the vein relative to the common carotid artery, and operator experience. The model was fitted with first-pass failure as the outcome and internally validated by bootstrap optimism correction with 1000 replications. Results: First-pass failure occurred in 46 of 120 procedures (38.3%). The odds of failure fell with each millimeter of anterior–posterior diameter (adjusted OR 0.83, 95% CI 0.72–0.95) and with each category of operator experience (adjusted OR 0.29, 95% CI 0.14–0.59), whereas the positional category showed no independent association. The optimism-corrected C-statistic was 0.743, with a calibration slope of 0.914 (95% CI 0.572–1.451) and a Brier score of 0.188. Conclusions: The ROTATE-ICU model estimates the probability of first-pass failure from three variables available at the bedside, of which vein diameter and operator experience carried most of the predictive weight. Because operator experience is itself one of the inputs, the model describes the risk of a given procedure performed by a given operator rather than a fixed property of the patient. The shrinkage-adjusted model and an open-access calculator are available; this is a preliminary model and external validation is now needed.