Introduction: Parenchymal hemorrhage (PH) is a feared complication of reperfusion therapies in patients with acute ischemic stroke (AIS), which can significantly worsen the prognosis. This study aimed to evaluate whether the MRI perfusion biomarkers could predict PH development in AIS patients undergoing reperfusion therapy. Methods: In this retrospective cohort study, AIS patients with anterior circulation large vessel occlusions included if the patients underwent pre-treatment diffusion-weighted imaging (DWI) and dynamic susceptibility contrast (DSC) perfusion MRI as well as follow up imaging to assess PH. Using a novel vascular model based on Bayesian framework, parametric maps including cerebral blood flow (CBF) and blood-brain barrier leakage were obtained. Apparent diffusion coefficient (ADC) from diffusion MRI was also included. Mask of baseline infarct (ADC < 620 × 10 -6 ) generated and registered to perfusion maps to extracted voxel values. The associations between PH and clinical and imaging parameters were assessed in univariate and multivariate analyses. A predictive model was developed using logistic regression analysis and a decision tree classifier following 5-fold stratified cross-validation. The performance of this model was tested in a separate external cohort of patients. Results: Among 124 patients included, 35 (28%) developed PH. Multivariate logistic regression analysis showed extreme values of perfusion parameters including 95 th% -leakage (OR=1.31, 95% CI: 1.1-1.63, p=0.006), 95 th% -CBF (OR=0.37, 95% CI: 0.13-0.89, p=0.002), and 5 th% -ADC (OR=0.98, 95% CI: 0.96-0.99, p<0.001) as independent biomarkers predictive of PH. The 5-fold cross validated combined model of these 3 variables (95 th% -leakage≥7.9, 95 th% -CBF≤0.7, and ADC≤360) resulted in an average AUC of 0.84 ± 0.05, 91% sensitivity and 78% specificity. In external validation cohort (n=20), the model correctly identified 80% of PHs with 90% specificity. Conclusions: We developed a multiparametric model by integrating baseline MRI biomarkers including ADC, CBF, and leakage to identify AIS patients at elevated risk of parenchymal hematoma after reperfusion. Extreme values of these markers were independently associated with PH and provided complementary information beyond single-parameter models. Incorporating multiparametric MRI into periprocedural workflows may enable more precise risk stratification and guide individualized treatment decisions.

Introduction: Venous transit delay (VTD), a marker of poor venous outflow, has been successfully measured on CTP using time-to-maximum (Tmax) maps and is associated with unfavorable outcomes in patients with acute ischemic stroke (AIS) caused by large vessel occlusion (LVO). Mean transit time (MTT) is a perfusion parameter that encompasses both arterial input and venous outflow allowing for a more robust representation of cerebral arteriovenous transit time compared to Tmax which is heavily arterially weighted. In this study, using a Bayesian framework, we aimed to compare the association of VTD calculated from MTT and arterial delay (Tmax equivalent in Bayesian) with poor clinical outcomes after reperfusion in AIS-LVO patients. Methods: This retrospective study included AIS-LVO patients with baseline CTP and follow-up MRI from two comprehensive stroke centers. CTPs were processed with the Bayesian method using an FDA approved software (Olea Medical) to generate MTT and arterial delay (AD) maps. VTD was measured in the transverse sinus on both the ischemic and normal side and the posterior superior sagittal sinus (SSS) in each patient. Two outcome measures were included: Infarct growth (poor: ≥ 10 mL) and functional outcome defined by 90-day modified Rankin Scale score (mRS) (poor: ≥3). Results: A total of 80 patients were included (51% infarct growth ≥10 mL, 34% 90-day mRS≥3). MTT-derived VTD from the transvers sinus ipsilateral to the ischemic hemisphere was significantly longer in patients with substantial infarct growth (p=0.0007) and poor functional outcome (p=0.001). Similarly, MTT-derived VTD from the SSS was significantly longer in patients with substantial infarct growth (p=0.003) and poor functional outcome (p=0.012). However, AD-derived VTDs from the transverse sinus nor SSS were not significantly different (p>0.05) for either outcome measure. The difference between VTDs measured on MTT and AD (ΔVTD MTT -VTD AD ) was significantly higher in patients with substantial infarct growth (p<0.0001 for transverse sinus and p=0.0002 for SSS) and poor functional outcomes (p=0.0002 for transverse sinus and p=0.0307 for SSS). Conclusion: MTT-derived VTD measured in the transverse sinus and SSS was significantly different between patients with poor and favorable outcomes, which was not reflected on arterially weighted AD maps. This signifies the potential for MTT to be superior for VTD measurements and prediction of outcomes in AIS patients.

BACKGROUND AND PURPOSE:Parenchymal hematoma (PH) is a severe complication of reperfusion therapy in acute ischemic stroke and is associated with poor functional outcome. We evaluated whether baseline quantitative imaging biomarkers derived from MR diffusion and perfusion could predict PH after reperfusion therapy. MATERIALS AND METHODS:In this retrospective study, consecutive adults with acute ischemic stroke due to large-vessel occlusion who underwent endovascular thrombectomy between 2015 and 2020 and had baseline MRI and follow-up imaging were included. Quantitative imaging biomarkers were extracted from the baseline ischemic core, including apparent diffusion coefficient (ADC), relative cerebral blood flow (rCBF), relative cerebral blood volume (rCBV), and relative K2 (rK2) derived from Bayesian-based DSC perfusion processing. The primary outcome was PH on follow-up imaging. Baseline imaging and clinical variables were compared between patients with and without PH. An elastic-net logistic regression model was developed using nested stratified group 5-fold cross-validation, and model performance was summarized by receiver operating characteristic analysis and SHAP-based feature importance. RESULTS:The final cohort included 100 patients, of whom 32 developed PH. Baseline demographic, clinical, and conventional imaging characteristics did not differ significantly between groups. Patients with PH had significantly lower values of rCBF, rCBV, and ADC and higher values of rK2 in comparison to those without PH. The final model incorporated 5th% ADC, rCBF, rCBV, and 95th% rK2 showed discrimination with an AUC of 0.88 ± 0.07 (mean ± SD) and overall accuracy of 77%. CONCLUSIONS:We developed a quantitative model using ADC, rCBV, rCBF, and rK2 from bassline MRI that can identify stroke patients with an increased risk of PH with approximately 77% accuracy.

Introduction: While mean transit time (MTT) was initially used as a surrogate for penumbra, today estimates rely on time-to-maximum (Tmax) due to poor image quality of MTT. Most commercially available software packages for CTP analysis use single value decomposition (SVD), which is highly sensitive to noise. The Bayesian approach is a robust probabilistic method with potential in improving image quality of perfusion maps. We aimed to perform a comparative analysis between MTT maps generated using Bayesian and SVD postprocessing to assess differences in diagnostic image quality and estimated penumbral volumes. Methods: This retrospective study included acute ischemic stroke patients with anterior large vessel occlusion stroke with baseline CTP from two comprehensive stroke centers. CTPs were processed using FDA-approved software: RAPID (iSchemaView)(SVD-1), Olea Medical (SVD-2) and the Bayesian method. Image quality assessment was performed by 2 independent board-certified neuroradiologists blinded to the method of postprocessing for the side of hypoperfusion (right, left) and overall image quality using a 1-4 Likert-like scale. Differences in image quality were tested by Mann-Whitney test and interobserver agreement was assessed by kappa statistics. For quantitative analysis, penumbral volumes were estimated by applying relative MTT> 1.4 to both SVD and Bayesian-generated MTT maps and compared against volumes obtained in routine clinical practice (Tmax>6 sec). Bland-Altman plots were generated for comparative analysis. Results: A total of 78 patients were included. The overall image quality was significantly (p<0.001) higher for Bayesian-MTT in comparison to SVD-MTT regardless of the commercial software used. There was substantial interobserver agreement for rating image quality scores for Bayesian-MTT (k=0.70), SVD-1 (k=0.69) and SVD-2 (k=0.78). Estimated penumbral volume (median, IQR) using Tmax>6 sec was 103, 67-131 mL, while MTT-estimated penumbral was 91, 62-127 mL for Bayesian and 25, 14-37 mL for SVD. The mean (SD) difference of estimated penumbral volume between Bayesian-MTT and Tmax was 3 (8) mL (p=0.60), while this difference was significantly (p<0.001) higher for SVD-MTT and Tmax (74, 10 mL). Discussion/Conclusion: MTT maps generated from a Bayesian framework significantly outperform SVD, which is used in most commercially available software packages, both in terms of image quality and estimation of penumbral volume.

BACKGROUND/OBJECTIVE:Arterial spin-labeling (ASL) MRI can measure perfusion signal adjacent to CSF spaces and may provide information regarding CSF-adjacent water transport physiology. We developed an automated pipeline to extract cortical-CSF interface (IF) perfusion for comparison between Alzheimer disease (AD) and cognitively normal controls. METHODS:In this retrospective study, participants with AD and cognitively normal controls with ASL and 3D T1-weighted MRI were included. Parenchymal cerebral blood flow (CBF) was evaluated globally (whole brain, gray matter, white matter) and across 12 brain regions. IF-perfusion values were quantified using an automated pipeline with tissue segmentation, registration, partial-volume-aware masking, and arterial-signal exclusion. Multivariable sensitivity analysis was performed, adjusting for age, sex, scanner vendor, and acquisition site. Predictive modeling was performed using LASSO-regularized logistic regression with nested repeated stratified cross-validation, and feature contributions were summarized using SHAP. RESULTS:A total of 51 patients with AD, age (median: 73 years) and 53 cognitively normal subjects (median age: 69) were included. Global CBF did not differ between groups. However, IF-perfusion values were significantly lower in AD (median 4.4, IQR 3.2-4.8) than in controls (median 5.2, IQR 4.5-6.3; p < 0.001). In multivariable analysis, AD remained independently associated with lower IF-perfusion after adjustment for age, sex, scanner vendor, and acquisition site (β = -1.88, 95% CI: -3.52 to -0.25; p = 0.025). In LASSO-regularized logistic regression model, only IF-perfusion and cuneus-CBF remained as significant contributing features for the prediction of AD probability with the highest contribution from IF. The final model showed mean ROC-AUC 0.81 in the training and AUC of 0.78 in the testing dataset. CONCLUSION:ASL-derived IF-perfusion is reduced in AD and may represent a promising imaging marker of altered CSF-adjacent water transport physiology. Further validation against established biomarkers of CSF dynamics and clearance pathways is warranted.

BACKGROUND:Artificial intelligence (AI) is increasingly integrated into medical imaging, but its adoption across neuroradiology departments remains uneven and poorly characterized. This State of Practice article reports the results of a national survey assessing how US neuroradiology departments are currently using AI - including which tools and applications are most common, how AI is perceived to affect workload and diagnostic performance, the nature of collaboration with AI vendors, and the primary barriers limiting broader adoption. METHODS:The survey was developed and distributed by the American Society of Neuroradiology (ASNR) Department Chair Working Group, a group of 18 US neuroradiologists serving as department chairs. A 19-item cross-sectional questionnaire - combining multiple-choice, multi-select, and open-ended items covering department demographics, AI usage and tools, clinical applications, perceived impact, vendor collaboration, barriers, pricing models, and future expectations - was distributed by e-mail between July 14 and August 11, 2025. Sixteen of 18 working group members (89%) completed the survey; responses were analyzed descriptively. KEY MESSAGE:AI use is already widespread among academic US neuroradiology departments (81%), concentrated heavily on stroke-related applications, yet most department chairs report that it has had minimal impact on workload so far and that tool performance remains inconsistent. Cost, integration challenges, and a lack of robust efficacy evidence remain the dominant barriers - underscoring that realizing AI's potential in neuroradiology will require closer collaboration between clinicians and vendors, rather than further tool proliferation alone.

Global imaging demand now exceeds capacity due to an aging population, rising comorbidities, expanded indications for imaging, and workforce shortages. These pressures are reshaping national health care policy. At the same time, the competing demands to expand access to imaging services while reducing costs are creating momentum to reimagine imaging delivery through the integration of transformative technologies like artificial intelligence (AI). This consensus statement from the International Society for Strategic Studies in Radiology (IS3R) is a call to action, outlining promising AI solutions that increase workflow efficiency and align radiology with health system and payer priorities. Reflecting expert consensus from the IS3R meeting in Dublin, Ireland, in August 2025, this statement defines strategic directions for radiology amid accelerating digitalization, rapid innovation, and data-driven patient care. This article covers four domains: image interpretation, capacity building, data governance, and systemic digital transformation. This article examines how AI can address rising imaging demand and costs while addressing challenges related to improving data quality, standardization, and interoperability. It also highlights the need to re-engineer workflows and adapt to new care models while maintaining high-quality services. Finally, it presents consensus recommendations outlining how AI can enable multimodal health data integration and advance alignment with value-based care.

Background In patients with acute ischemic stroke (AIS), ischemic core volume overestimation at diffusion-weighted MRI due to diffusion reversal has been described following successful reperfusion. Purpose To assess the potential role of the oxygen extraction fraction (OEF), derived from baseline dynamic susceptibility contrast (DSC) MRI, in diffusion reversal prediction in patients with AIS after successful reperfusion. Materials and Methods This retrospective study (January 2015 to December 2020) included patients with anterior-circulation large-vessel occlusion stroke who underwent pretreatment MRI, including diffusion-weighted imaging and DSC perfusion, achieved successful reperfusion (modified Thrombolysis in Cerebral Infarction score ≥2c), and underwent follow-up MRI within 48 hours. Ischemic core masks were generated from baseline MRI using an apparent diffusion coefficient (ADC) threshold of 620 × 10-6 mm2/sec or less, and infarct masks were generated from the final DWI scans. DSC perfusion was used to calculate the relative OEF (rOEF). Image co-registration and subtraction were performed to determine the final voxel fate as persistent infarct versus surviving tissue. Logistic regression and receiver operating characteristic (ROC) curve analyses were performed to assess imaging biomarker predictive value. Results Seventy patients were included (mean age, 72.2 years ± 13.1 [SD]; 40 female). Approximately 35% of the volume initially designated as the ischemic core showed diffusion reversal. Following logistic regression, the independent imaging variables associated with reducing the odds that ischemic core voxels would evolve to infarct voxels were ADC (odds ratio [OR], 0.99; 95% CI: 0.94, 1.0; P < .001) and rOEF (OR, 0.33; 95% CI: 0.32, 0.35; P < .001). A combined ADC plus rOEF model showed an area under the ROC curve (AUC) of 0.80 (95% CI: 0.80, 0.81), sensitivity of 74%, and specificity of 76% for infarct prediction, outperforming ADC alone (AUC, 0.65 [95% CI: 0.64, 0.66], sensitivity, 60%; specificity, 63%; P < .001). The added value of rOEF in delineating ischemic core was time-dependent, with the strongest negative impact within 120 minutes from stroke onset. Conclusion In patients with AIS for whom reperfusion was successful, the addition of rOEF to ADC improved ischemic core definition by predicting diffusion reversal and reducing core volume overestimation, especially in the hyperacute window. © RSNA, 2026 Supplemental material is available for this article. See also the editorial by Rapalino in this issue.

BACKGROUND AND PURPOSE:Artificial intelligence (AI) models have shown promise in neuroradiology, yet their real-world generalizability remains uncertain, partly due to variability in imaging acquisition and protocols. We aimed to evaluate the impact of data source, scanner manufacturer, scan mode, slice thickness, and the AI models, developed by participating teams, on AI performance in this secondary analysis of the 2019 American Society of Functional Neuroradiology (ASFNR) AI Competition. MATERIALS AND METHODS:We included 1177 anonymized noncontrast head CT scans from 5 institutions. Four teams participated, developing models to detect acute ischemic stroke, intracranial hemorrhage, and mass effect and to assess age-appropriate normality. Generalized estimating equations were used to evaluate the effects of the variables on model performance, and collinearity diagnostics were applied to exclude redundant variables. RESULTS:Due to collinearity with the scanner manufacturer, data source was excluded from the model. Across all tasks, the AI model significantly influenced the performance. The scanner manufacturer was significantly associated with accuracy in detecting intracranial hemorrhage and acute ischemic stroke but not mass effect or age-based normality. Slice thickness was significantly associated with detection of intracranial hemorrhage and mass effect, with thinner slices yielding higher accuracy, but it showed no effect on ischemic stroke or normality assessments. The scan mode did not significantly influence performance for any task. CONCLUSIONS:This secondary analysis demonstrates that imaging acquisition and protocol variability may significantly affect AI model performance. Scanner manufacturer, slice thickness, and the developed AI model were significantly associated with model accuracy, whereas scan mode had no significant impact. Among these, the developed AI model consistently proved the most influential, reflecting the importance of training data, model architecture, and preprocessing methods.

Background Clinical histories accompanying imaging orders guide protocol selection and diagnostic focus. However, they are often incomplete, potentially compromising diagnostic accuracy and workflow efficiency. Purpose To evaluate whether large language models (LLMs) can improve the clinical utility of provided imaging indications by leveraging clinical notes. Materials and Methods This retrospective study curated a dataset from deidentified electronic health records at the University of California San Francisco (January 2012 to August 2024), consisting of radiology reports with paired referring clinician-provided and radiologist-curated indications linked to clinical notes. The dataset was stratified across five body systems and five pathophysiologic categories to derive LLM selection and reader study internal test sets. For the reader study, 20 radiologists with 2-25 years of experience compared indications from the referring clinician, radiologist, and best-performing LLMs. Readers scored comprehensiveness, factuality, and conciseness and ranked indications for usefulness in protocoling, usefulness in interpretation, and overall ranking. Models and clinicians were compared using cumulative link mixed models with Tukey-adjusted post hoc comparisons. Results From 28 313 patients (mean age, 59 years ± 20.6 [SD]; 14 912 women), 250 examinations from 247 patients were sampled for the reader study. After nine exclusions, 241 examinations were analyzed, yielding 482 reader-examination evaluations. Indications from the best-performing proprietary (Claude 3.5 Sonnet; Anthropic) and open-source (Qwen 2.5-7B Instruct; Alibaba) LLM were rated as more comprehensive (Likert rating of 5: 37.14% and 28.42%, respectively; both P < .001) and factual (68.05% and 59.75%; both P < .001) than referring clinician indications. The proprietary LLM ranked most useful in protocoling (rank 1: 40.87%; all P < .001), useful in interpretation (44.61%; all P < .001), and overall ranking (44.19%, all P < .001). Comprehensiveness (65.77% of ratings; both P < .001) most strongly influenced overall rankings. Conclusion LLMs generated radiology-relevant indications from clinical notes that were more comprehensive and factual than clinician indications, and when generated by the proprietary LLM, were ranked most useful in protocoling and imaging interpretation. © RSNA, 2026 Supplemental material is available for this article. See also the editorial by Yilmaz and Cardoza-Ochoa in this issue.

BACKGROUND:In patients with acute ischemic stroke, infarct growth occurs despite successful reperfusion. Oxygen extraction fraction (OEF) has shown promising results in evaluating ischemic tissue viability and can now be quantified from routinely performed dynamic susceptibility contrast perfusion. We aimed to determine the association of OEF alterations within the ischemic tissue on pretreatment magnetic resonance imaging and infarct growth in patients who underwent successful reperfusion. METHODS:In this retrospective cohort study from the University of California, Los Angeles, between 2015 and 2020, patients were included if they had anterior circulation large vessel occlusion, achieved successful reperfusion (Thrombolysis in Cerebral Infarction ≥2b), had pretreatment dynamic susceptibility contrast perfusion and posttreatment magnetic resonance imaging within 48 hours from reperfusion. Dynamic susceptibility contrast-derived OEF values were quantified from the segmented ischemic core (apparent diffusion coefficient ≤620×10-6 mm2/s) and penumbra tissue (time-to-maximum [Tmax] >6 s) on pretreatment magnetic resonance imaging and normalized to contralateral hemisphere (relative oxygen extraction fraction [OEFr]). Primary outcome was substantial infarct growth ≥10 mL, and secondary outcomes were continuous measures of infarct growth volume and penumbra-to-infarct conversion ratio. The associations between baseline clinical and imaging variables, including OEFr and outcome measures, were tested by multivariate and regression analysis. RESULTS:Among 89 patients who met inclusion criteria, 33 (37%) patients had infarct growth ≥10 mL. Patients with infarct growth had significantly (P<0.0001) lower penumbra-OEFr values compared with those without infarct growth. There was significant association between penumbra OEFr and infarct growth (β=-2.9 [95% CI, -5.0 to -0.8]; P=0.007) and similarly for penumbra-to-infarct conversion ratio (β=-10.4 [95% CI, -19.6 to -1.2]; P=0.028). CONCLUSIONS:Our results showed penumbra-OEFr is a promising imaging biomarker for predicting infarct growth in acute ischemic stroke following successful reperfusion. Although elevation of penumbra-OEFr is protective, patients with lower penumbra-OEFr values sustained further ischemic injury and infarct growth.

Computational competitions are the standard for benchmarking medical image analysis algorithms, but they typically use small curated test datasets acquired at a few centers, leaving a gap to the reality of diverse multicentric patient data. To this end, the Federated Tumor Segmentation (FeTS) Challenge represents the paradigm for real-world algorithmic performance evaluation. The FeTS challenge is a competition to benchmark (i) federated learning aggregation algorithms and (ii) state-of-the-art segmentation algorithms, across multiple international sites. Weight aggregation and client selection techniques were compared using a multicentric brain tumor dataset in realistic federated learning simulations, yielding benefits for adaptive weight aggregation, and efficiency gains through client sampling. Quantitative performance evaluation of state-of-the-art segmentation algorithms on data distributed internationally across 32 institutions yielded good generalization on average, albeit the worst-case performance revealed data-specific modes of failure. Similar multi-site setups can help validate the real-world utility of healthcare AI algorithms in the future.

Brain metastases are a frequent and debilitating manifestation of advanced cancer. Here, we collect and analyze neuroimaging of 3,065 cancer patients with 13,067 brain metastases, representing an extensive collection for research. We find that metastases predominantly localize to high perfusion areas near the grey-white matter junction, but also identify notable differences depending on the primary cancer histology as well as brain regions which do not conform to this relationship. Lung and breast cancers, in contrast to melanoma, frequently metastasize to the cerebellum, hinting at biological pathways of spread. Additionally, the deep brain structures are relatively spared from metastasis, regardless of primary cancer type. Leveraging this data, we propose a probabilistic brain metastasis risk model to enhance the therapeutic ratio of whole-brain radiotherapy by targeting high risk areas while preserving cortical and subcortical brain regions of functional significance and low metastasis risk, potentially reducing the cognitive side effects of therapy.

This study shows that MRI acceleration techniques reduce energy use, carbon emissions, and costs, while increasing imaging capacity and practice revenue, offering a sustainable and efficient strategy for health care operations.

Background:There are no disease modifying therapies for Huntington's disease (HD), a rare but fatal genetic neurodegenerative condition. To develop and test new management strategies, a better understanding of the mechanisms underlying HD progression is needed. Aberrant changes in thalamo-cortical and striato-cerebellar circuitry have been observed in asymptomatic HD, along with transient enlargement of the dentate nucleus. Purpose:To evaluate the relationship between thalamo-cerebellar connectivity and HD progression. Study Type:Prospective and retrospective. Population:Patients with HD and healthy controls from a single-center dataset (n=34), and patients from the public TRACK-HD dataset (n=91). Field strength/Sequence:3T and 7T. Assessment:Thalamo-cerebellar connectivity was compared across patients and controls and related to motor scores and predicted years to symptom onset. Cross-sectional findings were validated within-patient by mapping changes in individual connectivity over time. HD effects on cognitive performance were also explored and related to connectivity. Statistical Tests:Kruskal-Wallis with post hoc Dunn's tests and Pearson correlations (p significant <0.05). Results:In the 7T cohort, significant premanifest and control group differences in thalamo-dentate connectivity were observed (p Dunn <0.05, η 2 =.19-.22), with manifest HD connectivity approaching normative values. Thalamic connectivity with the dentate nucleus and anterior cerebellum also correlated with years to onset (p Den =0.06, r=0.42, p Ant <0.05, r=-0.45), together indicating potential transient functional alterations in premanifest HD. Similar patterns were observed between connectivity (thalamus to dentate nucleus and anterior lobe) and cognitive performance scores across all subjects (p<0.05, r Den =-0.17, r Ant =-0.18). In the premanifest TRACK-HD cohort, connectivity of multiple thalamo-cerebellar connections correlated with years to onset, revealing distinct patterns for patients with low versus high motor scores, again indicative of potential transient alterations. Exploratory non-parametric regression of serial imaging data further supported these findings. Data conclusion:Transient changes in thalamo-cerebellar connectivity are seen in premanifest HD with increasing progression. More studies are needed to validate this potentially useful biomarker.

Importance:Annual rates of first intracranial hemorrhage (ICH) from unruptured brain arteriovenous malformations (AVMs) are often quoted as 2% to 4% in clinical practice. Precise estimates and risk factors are unavailable to inform treatment decisions. Objective:To provide estimates of rates and risk factors for first ICH in a large cohort study of unruptured brain AVMs. Design, Setting, and Participants:The Multicenter Arteriovenous Malformation Research Study (MARS) included data from 9 cohorts, each contributing 100 or more unruptured brain AVMs. The study was conducted from 2017 to 2023 with retrospective and prospective data collection for existing cohorts and/or new recruitment. This was an international study (2 population-based and 7 referral-based cohorts) that included participants diagnosed with an unruptured brain AVM. Exposures:Demographic, clinical, and angiographic characteristics. Main Outcome and Measure:The primary outcome was time to first ICH after diagnosis of unruptured brain AVM. Data were collected using standardized definitions; missing data were imputed. Cox regression analysis was performed, censoring at first brain AVM treatment, death, or last visit, and allowing baseline hazards to vary by cohort. Results:A total of 3225 individuals were eligible for participation in this study. After 195 exclusions, 3030 participants (median [IQR] age, 38 [25-50] years; 1524 female [50.3%]) were included. Among 2989 participants with unruptured brain AVMs, 1333 (45%) presented with seizure. The median (IQR) maximal brain AVM diameter was 3.1 (2.2-4.4) cm, 248 of 2466 AVMs (10%) had exclusively deep venous drainage, 297 of 2690 (11%) were in supratentorial deep or cerebellar locations, and 457 of 2440 (19%) had associated arterial aneurysms. First ICH occurred in 159 participants over 11 339 person-years of follow-up for an ICH rate of 1.40 (95% CI, 1.20-1.64) per 100 person-years. Significant independent risk factors included (1) increasing age category at diagnosis (hazard ratio [HR], 0.87; 95% CI, 0.53-1.41 for those aged 20 to 39 years; HR, 1.23; 95% CI, 0.74-2.04 for those aged 40 to 59 years; and HR, 2.01; 95% CI, 1.14-3.57 for those aged 60 years vs younger than 20 years; P = .008), (2) presence of associated aneurysms (HR, 1.66; 95% CI, 1.06-2.59; P = .03), and (3) cerebellar or supratentorial deep location (HR, 1.87; 95% CI, 1.16-3.00; P = .01). Conclusions and Relevance:The annual ICH rate from unruptured brain AVM was lower than that commonly cited in clinical practice. Increasing age, associated arterial aneurysms, and cerebellar or supratentorial deep brain AVM location were associated with risk of first ICH. These results may be used to counsel patients about the natural history of unruptured brain AVMs.