Introduction: Hospitalization anemia is common in intracerebral hemorrhage (ICH) and associates with poor long-term outcomes. In other critical illnesses, hospitalization anemia is similarly common, yet can persist months after discharge and impacts clinical outcomes. In ICH, prevalence of post-hospitalization anemia is poorly characterized and its relationship with outcomes is unknown. We investigated relationships of 3-month post discharge hemoglobin levels with long-term ICH outcomes. Methods: Spontaneous ICH patients enrolled into a single-center, prospective observational study between 2009 and 2019 with available hospitalization and 3-month post-discharge complete blood count (CBC) assessments were assessed. Poor 6 and 12-month neurological outcomes (Modified Rankin Scale 4–6) were assessed. Multivariable logistic regression models assessed relationships between post discharge hemoglobin and poor neurological outcomes, adjusting for ICH severity and baseline demographics. Results: Of the initial cohort of 810 ICH patients, 67% of patients had anemia present on hospital discharge and 221 of these patients had follow-up CBC testing at 3-months post discharge. Amongst this analyzed cohort of 221 ICH patients with post discharge CBC testing, 13% met criteria for anemia. In our regression analyses, we identified that lower post discharge hemoglobin associated with increased odds of poor 6-month outcomes (adjusted OR 0.65, 95% CI: 0.50-0.84, p=0.001) and 12-month outcomes (adjusted OR 0.68, 95% CI: 0.52-0.88, p=0.004). Repeated measure analyses of serial hospitalization CBC and post discharge CBC data revealed significant differences in hemoglobin recoveries over time amongst patients with and without poor long-term outcomes. Conclusions: Post-discharge anemia remains prevalent after ICH and relates to poor long-term outcomes. Further work is needed to clarify the generalizability of these findings and to define factors contributing to anemia recovery after ICH and critical illness. This may elucidate hospitalization or post-discharge treatment targets and management strategies to improve ICH outcomes.

The reliability of electronic medical records (EMR) on exposure to sedative and analgesic medications in patients with acute disorders of consciousness is unknown. Our objective was to quantify the accuracy of sedative and analgesic infusion rates derived from the EMR to support its use in Big Data clinical research. We conducted a secondary analysis of prospective cohort studies enrolling patients with critical illness who were unresponsive to verbal commands after acute brain injury. During standardized behavioral assessments, research coordinators documented infusing sedative and analgesic medications in case report forms (CRF; reference standard). We paired infusion rates from the CRF with corresponding infusion rates from the EMR (index). For each drug with ≥ 10 EMR-CRF infusion rate pairs, we calculated the concordance correlation coefficient and created Bland–Altman plots to estimate biases and limits of agreement (LOA). Among 63 included patients (median [interquartile range (IQR)] age: 61 [46–72] years; 22 [35

Objective: Intracranial pressure (ICP) waveform morphology reflects brain compliance and cerebrospinal fluid dynamics. Existing monitoring methods fail to fully capture complex temporal patterns nor enable real-time interactive analysis to improve clinical decision-making. Methods: We trained a transformer-based foundation model to capture temporal dynamics and generate embeddings from physiological data. The model was fine-tuned on physiological waveform data from patients with intracerebral hemorrhage (ICH) admitted to Columbia University Irving Medical Center (CUIMC) and validated on two non-overlapping datasets: a) patients with cerebral external ventricular drainage (EVD) at CUIMC and b) a synthetic ICP dataset. The embeddings generated from the foundation model were used to train a support vector machine (SVM) classifier to classify different morphologies. Model performance was evaluated using area under the receiver operating curve (AUC) and confusion matrices, by splitting the dataset into training and testing. We developed a graphical user interface to enable ad-hoc analysis, fine-tune models, and visualize ICP trends. Results: A total of 190 patients between March 2009 and August 2013 with ICH were included to fine-tune the foundation model with a median length of stay (LOS) of 5 [4-9.25] days, and a Glasgow Coma Scale (GCS) score of 11 [7-15]. 6s2 of these patients had ICP waveform data. A total of 23 other patients (train: 11, test 12) from January 2021 to August 2023 with EVD were used to train the SVM model to classify different ICP morphologies; with median LOS 17 [12-23] days, and GCS 7 [5 13]. Two trained experts (YL, GG) labeled 8406 ICP (train:3613, test: 4793) pulses. The model achieved AUCs of 0.90 for 3-peak compliant, 0.93 for single-peak non-compliant, and 0.78 for multi peak non-compliant waveforms. On simulated ICP data, the AUCs were 1.00 for all the waveform classes. However, the confusion matrix analysis revealed that 1-peak compliant waveforms were classified with 77.5% accuracy while all other categories had 100% accuracy. Conclusion: A deep learning-based foundation model optimized for analyzing invasive ICP waveforms can extract clinically relevant information about cerebral compliance. The performance was best for distinguishing compliant from non-compliant waveforms.

OBJECTIVE:Impairment of consciousness is frequent in patients with supratentorial brain injury, but underlying mechanisms are poorly understood. Large-scale network dysfunction has been suggested as a unifying concept for the diverse focal infra- and supratentorial brain lesions associated with unconsciousness. Functional imaging studies have revealed that lesions in different locations that cause the same symptom can be linked to common networks using lesion network mapping (LNM). METHODS:To identify a commonly injured network for impaired consciousness, LNM was used in a prospective, bi-center observational cohort of 115 humans suffering from focal brain injury due to supratentorial intracerebral hemorrhage (ICH). RESULTS:Contrasting patients with preserved (n = 45) and impaired consciousness (n = 70), we identified a common network of brain regions more likely to be connected to focal brain lesions associated with impaired consciousness. Lesions in unconscious patients were more likely connected to a network including fronto-insulo-limbic association cortex, subcortical arousal and neuromodulatory nodes, and the ventral striato-pallidal circuitry. Data suggested preferential connectivity to a combination of arousal circuitry and association cortices. INTERPRETATION:Patients with impaired consciousness shared a common pattern of distributed network involvement that we infer to be mechanistically relevant to the regulation of consciousness that may serve as promising targets of neuromodulatory therapies. ANN NEUROL 2026;100:206-221.

Purpose: Free-text clinical notes contain rich prognostic information often lost in traditional models limited to structured variables (e.g., age, sex, NIHSS). Deep learning–based natural language processing (NLP) can leverage this information without manual variable extraction and may uncover risk factors beyond established predictors. We evaluated different NLP strategies for predicting 90-day mortality from ICU notes across multiple stroke types and examined model explainability using SHapley Additive exPlanations (SHAP) value quantification. Methods: We used the Medical Information Mart for Intensive Care (MIMIC-IV) database (>40,000 ICU patients; Beth Israel Deaconess Medical Center, 2008–2019) and identified 7,511 patients with acute ischemic stroke, spontaneous intracerebral hemorrhage (ICH), non-traumatic subarachnoid hemorrhage (SAH), and traumatic SAH. We compared four NLP strategies using transformer-based models designed to process varying text lengths: (1) BioBERT with full free-text notes (512 tokens), (2) BioBERT with keyword-focused 512-token summaries, (3) Longformer (4,096 tokens), and (4) dual-stream BioBERT combining full notes and summaries. Separate models were trained for each stroke subtype. SHAP quantified the contribution of individual text tokens to patient-level risk predictions. Results: Longformer and dual-stream BioBERT consistently outperformed other approaches. In 5-fold cross-validation, best-performing models achieved mean AUCs of 0.83±0.06 (ischemic stroke, n=1,819), 0.81±0.08 (ICH, n=1,657), 0.85±0.04 (non-traumatic SAH, n=877), and 0.82±0.02 (traumatic SAH, n=3,158), summarized in Table 1. SHAP identified high-impact tokens from free text such as “intubation,” “transfer,” specific comorbidities, and medications, many of which extend beyond known prognostic variables (Fig 1). Conclusion: Transformer-based NLP models, particularly those handling longer text sequences or combining full and focused inputs, can accurately predict 90-day mortality from free-text ICU notes across stroke types. SHAP explainability highlights novel high-risk features, suggesting a potential role for automated, real-time risk stratification directly from the electronic health record to guide early intervention. Such free text-based deep learning models can accelerate risk stratification model development by bypassing variable extraction and potentially identifying risk factors beyond known predictors.

Cerebral blood flow is essential for brain function and is governed by cerebral perfusion pressure and physiological mechanisms collectively referred to as cerebral vascular regulation (CVR). Direct measurement of cerebral blood flow and individual CVR mechanisms is challenging and often unavailable clinically, limiting personalization of blood flow and CVR targets, particularly in neurological injury. Physics-informed digital twins enable estimation, tracking, and forecasting of unmeasured physiological states from limited data. Therefore, a digital twin of cerebral hemodynamics that includes CVR mechanisms could help overcome measurement barriers. Here, we introduce CereBRLSIM (Cerebral Blood Regulation Latent State Inference and Modeling), a digital twin that assimilates physiological knowledge and patient data to infer CVR function and predict intracranial hemodynamics. Using in vivo experiments and simulated data, CereBRLSIM predicted cerebral blood flow and estimated myogenic, endothelial, and metabolic CVR mechanism dynamics. When personalized to data from six neurocritical care patients, CereBRLSIM differentiated cerebral hemodynamic pressure-flow phenotypes, predicted outcomes, and forecasted blood flow with higher accuracy than machine learning models. This work provides an interpretable and clinically compatible approach for quantifying CVR function and forecasting cerebral blood flow, potentially enabling precision diagnostics and understanding cerebral hemodynamics.

ABSTRACT Background The Functional Outcome in Patients with Primary Intracerebral Hemorrhage (FUNC) score was initially validated for prediction of functional independence on the Glasgow Outcome Scale (GOS) 90 days after intracerebral hemorrhage (ICH), but recovery often extends beyond three months. Aims Our objective was to extend the FUNC score for prediction of 12-month functional independence to strengthen its utility for family counseling and research methodology. Methods We conducted a single-center prospective cohort study enrolling adult patients with primary ICH between February 2009 and January 2018. We calculated FUNC scores at admission and assessed GOS 12 months after ICH. The primary outcome was 12-month functional independence, defined as a GOS score ≥4. We calculated the area under the receiver operating characteristic curve (AUC) of the FUNC score using logistic regression, handling missing GOS with multiple imputation by chained equations. We evaluated score calibration using a calibration curve and the Brier score, and we assessed clinical utility using decision curve analysis. We explored the statistical efficiency gains of using FUNC-based sliding dichotomy thresholds for favorable outcome definitions by running simulations of a clinical trial with 1:1 randomization. We ran 5000 simulations for each sample size (100 to 1000, in increments of 10) and treatment effect (odds ratio of 1.5, 2.0 and 2.5) combination and calculated efficiency gains for each respective treatment effect as the percentage reduction in sample size required to have 80% power using sliding versus fixed dichotomy thresholds. Results A total of 535 patients were included (median [IQR] age 68 [54-79], 237 [44%] female, median [IQR] NIHSS 16 [6-25], median [IQR] FUNC 8 [6-9]). Overall, 99 of 445 (22%) patients with known 12-month GOS achieved functional independence. The FUNC score had an AUC of 0.79 (95%-CI: 0.75-0.84) for 12-month functional independence. The calibration plot was reasonable, with modest evidence of overestimation at low predicted probabilities, and the Brier score was 0.15. A net benefit was observed across 5-50% threshold probabilities. Sliding dichotomy had an efficiency gain of 27% for a treatment effect of OR=2.0, and a gain of 22% for a treatment effect of OR=2.5. The efficiency gain for a treatment effect of OR=1.5 could not be calculated because the fixed dichotomy did not reach 80% power despite a sample size of 1000 patients. Conclusions The FUNC score’s predictive performance for 12-month functional independence was comparable to its originally validated 3-month discrimination. Following external validation across centers, the FUNC score may be leveraged to counsel families on global measures of long-term functional independence and to implement sliding dichotomy methodology in ICH research.

Atrial fibrillation detected after stroke (AFDAS) refers to newly identified atrial fibrillation occurring after an index cerebrovascular event, most commonly ischemic stroke, but little is known about its occurrence and clinical implications in patients with spontaneous intracerebral hemorrhage (ICH). We aimed to characterize AFDAS risk factors in ICH and AFDAS’ relationship with long-term ICH outcomes. We retrospectively analyzed consecutively admitted and enrolled patients with spontaneous ICH to a single tertiary comprehensive stroke center between 2009 and 2018 who underwent continuous cardiac monitoring. Patients were categorized as having sinus rhythm [SR (reference)], atrial fibrillation detected after stroke (AFDAS, defined here as atrial fibrillation (AF) detected during the index ICH hospitalization), and known AF. Demographic, clinical, imaging, and hospitalization complication data were compared across rhythm groups. Poor modified Rankin scale (mRS 4–6) at 3-month follow-up was assessed as the primary clinical outcome. Multivariable logistic regression assessed independent relationships of AFDAS with poor outcomes adjusting for demographics, ICH severity, and AFDAS risk factors. Amongst 532 patients with ICH, AFDAS occurred in 5

PURPOSE:Cognitive motor dissociation (CMD) is associated with long-term recovery in acute brain injury, but CMD testing is only available in few centers. Our objective was to identify surface EEG patterns with high sensitivity or positive predictive value (PPV) for CMD in patients with acute disorders of consciousness to refine allocation of this resource-intensive test. METHODS:In this observational cohort study, we enrolled clinically unresponsive, acutely brain injured patients who underwent continuous surface EEG and CMD assessments. CMD was detected by applying a machine learning algorithm to EEG acquired during a motor command paradigm presentation. Electroencephalographers blinded to CMD test results applied standardized ACNS criteria to the EEGs acquired during CMD assessments. We calculated accuracy measures of surface EEG findings for CMD test results using generalized estimating equations, with an exchangeable matrix and accounting for repeated measures per patient. RESULTS:We included 185 patients (mean age: 62 ± 17; 85 [46%] female) and 282 CMD assessments. CMD testing was positive in 39 (14%) assessments. Sensitivity and PPV of normal background voltage, symmetry, and continuity were, respectively, 77% (95% CI: 60%-88%) and 19% (95% CI: 13%-26%), 74% (95% CI: 58%-86%) and 14% (95% CI: 10%-20%), and 74% (95% CI: 58%-86%) and 14% (95% CI: 9%-19%). All EEGs with burst suppression, suppression, sporadic epileptiform discharges, lateralized periodic discharges, bilateral independent periodic discharges, electrographic seizures, and brief potentially ictal rhythmic discharges had negative CMD tests. CONCLUSIONS:Surface EEG findings are not reliable to screen for CMD or to identify patterns conferring higher CMD pretest probability.

Background and Purpose: Diffusion weighted imaging (DWI) ischemic lesions identified after intracerebral hemorrhage (ICH) are known to associate with poor outcomes. These lesions may be attributable to microthrombosis, yet it is unclear whether specific thrombotic risk factors for ICH patients contribute to increased DWI lesion risk. We sought to assess whether increased platelet counts associate with DWI lesions after ICH. Methods: Hospitalized acute spontaneous ICH patients with baseline platelet counts and inpatient magnetic resonance imaging (MRI) were assessed from a single-center ICH cohort and a separate external, multi-center validation ICH cohort. Patients with systemic coagulopathy unrelated to platelet counts (PT>20, PTT>50, INR >1.7) were excluded. Baseline platelet count was assessed as a continuous exposure variable. DWI ischemic lesion presence was defined as the primary outcome. The relationship of platelet counts with DWI lesion presence was assessed using multivariable logistic regression models adjusting for baseline characteristics, ICH score, time between admission and MRI, and change in systolic blood pressure within first 24 hours of admission. Secondary analyses were performed stratified by ICH location in the validation cohort. Results: We identified 184 and 874 ICH patients for analyses from the primary and external validation cohorts, respectively. Mean baseline platelet count was 226.2 and 229.7 and DWI lesions were present in 30% and 27% of patients from each cohort, respectively. In our primary single-center cohort, higher platelet counts significantly associated with DWI lesions (adjusted OR per 10,000 platelets/uL 1.063 [95% CI 1.002-1.127]). We identified similar relationships in our multi-center validation cohort (adjusted OR 1.029 [1.006-1.052]). Stratified analyses by ICH location revealed that these relationships were again seen in lobar ICH patients (adjusted OR 1.041 [1.004-1.079]) but not deep ICH (adjusted OR 1.020 [0.991-1.050]). Conclusion: Higher baseline platelet counts are associated with ischemic lesions in spontaneous ICH patients, with lobar ICH patients more vulnerable to this relationship. Given ongoing work to assess DWI lesion prevention with antithrombotic medications, further work is required to clarify whether these findings can help target DWI lesion prevention and treatment strategies.

A survey of 44 worldwide clinical experts in continuous monitoring of cerebral autoregulation (CCA) revealed that while half used a CCA index in clinical decision making, only 39

BACKGROUND: Disturbed cerebral autoregulation remains a theoretical contributor to posterior reversible encephalopathy syndrome (PRES), but it has not been captured before. We report invasively measured autoregulation indices in a 50-year-old female with aneurysmal subarachnoid hemorrhage. CASE SUMMARY: PRES developed after induced hypertension and intra-arterial nimodipine infusion as treatment for delayed cerebral ischemia. The poorest autoregulation values were observed a week before PRES, at the time of delayed cerebral ischemia and sepsis. A second, less pronounced deterioration of autoregulation occurred before PRES diagnosis. Relevant comorbidities potentially related to PRES were fetal-type arterial supply to the posterior circulation, pneumonia, and steroid application. CONCLUSIONS: This is the first reported case with cerebral autoregulation measurements during PRES. Hypothetically, autoregulation could have contributed to PRES in a setting of persistent hypertension, immunological factors, and a more susceptible posterior circulation. Further data are required to reliably investigate the relationship between autoregulation and PRES.

BACKGROUND AND PURPOSE:Robustness against input data perturbations is essential for deploying deep learning models in clinical practice. Adversarial attacks involve subtle, voxel-level manipulations of scans to increase deep learning models' prediction errors. Testing deep learning model performance on examples of adversarial images provides a measure of robustness, and including adversarial images in the training set can improve the model's robustness. In this study, we examined adversarial training and input modifications to improve the robustness of deep learning models in predicting hematoma expansion (HE) from admission head CTs of patients with acute intracerebral hemorrhage (ICH). MATERIALS AND METHODS:We used a multicenter cohort of n = 890 patients for cross-validation/training, and a cohort of n = 684 consecutive patients with ICH from 2 stroke centers for independent validation. Fast gradient sign method (FGSM) and projected gradient descent (PGD) adversarial attacks were applied for training and testing. We developed and tested 4 different models to predict ≥3 mL, ≥6 mL, ≥9 mL, and ≥12 mL HE in an independent validation cohort applying receiver operating characteristics area under the curve (AUC). We examined varying mixtures of adversarial and nonperturbed (clean) scans for training as well as including additional input from the hyperparameter-free Otsu multithreshold segmentation for model. RESULTS:When deep learning models trained solely on clean scans were tested with PGD and FGSM adversarial images, the average HE prediction AUC decreased from 0.8 to 0.67 and 0.71, respectively. Overall, the best performing strategy to improve model robustness was training with 5:3 mix of clean and PGD adversarial scans and addition of Otsu multithreshold segmentation to model input, increasing the average AUC to 0.77 against both PGD and FGSM adversarial attacks. Adversarial training with FGSM improved robustness against similar type attack but offered limited cross-attack robustness against PGD-type images. CONCLUSIONS:Adversarial training and inclusion of threshold-based segmentation as an additional input can improve deep learning model robustness in prediction of HE from admission head CTs in acute ICH.

Acute ischemic lesions seen on brain magnetic resonance imaging (MRI) are associated with poor intracerebral hemorrhage (ICH) outcomes, but drivers for these lesions are unknown. Rapid hemoglobin decrements occur in the initial days after ICH and may impair brain oxygen delivery. We investigated whether acute hemoglobin decrements after ICH are associated with MRI ischemic lesions and poor long-term ICH outcomes. Consecutive patients with acute spontaneous ICH enrolled into a single-center prospective cohort study were assessed. Change in hemoglobin levels from admission to brain MRI was defined as the exposure variable. The presence of MRI ischemic lesions on diffusion-weighted imaging was the primary radiographic outcome. Poor 6-month modified Rankin Scale score (4–6) was assessed as our clinical outcome. Separate regression models assessed relationships between exposure and outcomes adjusting for relevant confounders. These relationships were also assessed in a separate prospective single-center cohort of patients with ICH receiving minimally invasive hematoma evacuation. Of 190 patients analyzed in our primary cohort, the mean age was 66.7 years, the baseline hemoglobin level was 13.4 g/dL, and 32

Artificial Intelligence (AI) is rapidly transforming the landscape of critical care, offering opportunities for enhanced diagnostic precision and personalized patient management. However, its integration into ICU clinical practice presents significant challenges related to equity, transparency, and the patient-clinician relationship. To address these concerns, a multidisciplinary team of experts was established to assess the current state and future trajectory of AI in critical care. This consensus identified key challenges and proposed actionable recommendations to guide AI implementation in this high-stakes field. Here we present a call to action for the critical care community, to bridge the gap between AI advancements and the need for humanized, patient-centred care. Our goal is to ensure a smooth transition to personalized medicine while, (1) maintaining equitable and unbiased decision-making, (2) fostering the development of a collaborative research network across ICUs, emergency departments, and operating rooms to promote data sharing and harmonization, and (3) addressing the necessary educational and regulatory shifts required for responsible AI deployment. AI integration into critical care demands coordinated efforts among clinicians, patients, industry leaders, and regulators to ensure patient safety and maximize societal benefit. The recommendations outlined here provide a foundation for the ethical and effective implementation of AI in critical care medicine.