Document resource
Introduction Ruptured intracranial aneurysms are a significant cause of morbidity and mortality. While individual-level risk factors for rupture are well established, the influence of meteorological variables has been routinely opined upon but remains unclear.Methods We retrospectively analyzed 1,504 patients with confirmed intracranial aneurysms from two comprehensive stroke centers in Upstate New York (2018-2024). Daily weather data were matched to the date of presentation. Multivariable logistic regression was performed after variance inflation factor analysis to remove collinear features. We additionally trained extreme gradient boosting (XGBoost) models using weather alone and in combination with demographics (age, sex, race, smoking, family history). Model performance was assessed using AUC, F2 score, and SHapley Additive exPlanations (SHAP).Results Among 1,504 patients, 377 (25.1%) presented with ruptured aneurysms. VIF analysis excluded 9 collinear weather variables. Univariate logistic regression identified four significant predictors of rupture: lower humidity (OR 0.671, 95% CI 0.471-0.955, p = 0.0133), greater snow depth (OR 6.616, 95% CI 5.028-110,950, p = 0.0096), lower ultraviolet index (OR 0.931, 95% CI 0.187-0.831, p = 0.0144), and moon phase (OR 8.660, 95% CI 3.217-1.033e+07, p = 0.0235). On multivariable analysis, only female sex remained protective (OR 0.605, 95% CI 0.386-0.949, p = 0.0285), while sea-level pressure trended toward significance (OR 0.962, p = 0.0523). Logistic model sensitivity was 3.97%. The XGBoost weather-only model yielded an AUC of 0.560 and a recall of 78%. Adding demographics improved AUC to 0.590 and recall to 83%, with SHAP identifying family history, precipitation, and cloud cover as key features.Conclusions While several weather variables correlated with rupture risk in univariate analysis, the overall predictive value was limited. Machine learning improved sensitivity but confirmed patient-level features as dominant contributors. Weather data alone may have limited clinical utility in rupture prediction within temperate regions.SHAP Summary Plot. This plot ranks all input variables by their overall impact on model predictions. Each horizontal strip represents one variable, with dots showing individual patient contributions. The x-axis is the SHAP value (positive = increased rupture prediction; negative = decreased prediction). The color gradient reflects the original value of the feature: red for high and blue for low. For example, red dots on the left side for ‘cloudcover’ suggest that high cloud cover values generally decrease rupture risk, while blue dots on the right for ‘AGE’ suggest that lower age tended to push the model towards predicting rupture.Disclosures A. Goyal: None. A. Gajjar: None. A. Naqvi: None. S. Bheemireddy: None. A. Custozzo: None. V. Jaikumar: None. A. Siddiqui: None. A. Boulos: None. J. Dalfino: None. A. Paul: None.Abstract O-066 Figure 1