Abstract WP343: Development of AneuScreen+ ^TM , a First-of-its-Kind Blood-Blood Diagnostic for Intracranial Aneurysm

Introduction: Intracranial aneurysm (IA) affects 3-5% of US population, and can lead to subarachnoid hemorrhage with high rates of morbidity and mortality. It is crucial to assess rupture risk so that clinicians can treat those at high risk of rupture and avoid risky surgeries for those with low risk. We are developing a low-cost, non-invasive blood RNA biomarker for aneurysm called AneuScreen TM . Methods: We performed feature selection and preliminary model development using an existing dataset of 191 transcriptomes from individuals with IA to identify genes associated with IA rupture risk, determined by the PHASES score. We considered genes selected by differential expression analysis (edgeR), minimum redundancy-maximum relevance algorithm over 500 randomizations, and our previous studies. For the selected genes, we conducted analytical testing using pooled control human blood to ensure the designed probes met the signaling requirements. We subsequently tested the assay on a large group of clinical blood samples, collected under IRB approval from 4 US centers. The resultant expression data along with clinical data was used to refine the algorithm via a robust machine learning pipeline. Multiple feature selection strategies were applied independently across 500 stratified, random splits (80/20 train-test). Features selected in at least 50% of iterations were deemed stable. These features were then used to train and evaluate four classifiers, each optimized via 5-fold cross-validation in 500 randomizations in the training set. Model performance was assessed using ROC AUC, sensitivity, specificity, and accuracy. SHAP analysis was used to interpret model decisions based on the stable feature set. Results: We identified 21 genes associated with aneurysm risk. A logistic regression model using expression levels achieved an AUC>0.8. The designed PCR probes met all analytical testing requirements, detectability, consistency across replicates, and linearity. After assay testing and model optimization, the best model was a logistic regression model that demonstrated high discriminative ability, with a mean ROC AUC of 0.83 and sensitivity>0.77 in testing. SHAP analysis reflects both genes and clinical features are important. Conclusion: The AneuScreen TM risk assessment module was tested in a large cohort of individuals with IA and could predict high-risk cases with an AUC of 0.83, demonstrating a potential for using circulating blood to assess IA risk.