Abstract 172: A Machine Learning Model for Prediction of Short‐Term Functional Outcomes in Cerebral Venous Thrombosis

Background Dural Sinus and Cerebral Vein Thrombosis (CVT) represents a rare yet significant cause of stroke, with distinct pathophysiological mechanisms compared to arterial strokes. Despite the progress in machine learning‐based predictive models for arterial stroke, which have significantly advanced prognostication and treatment strategies, there remains a notable absence of similar models for CVT. Methods A single‐center, retrospective cohort study of adult patients diagnosed with acute CVT confirmed on contrast‐enhanced CT/MR Imaging at admission was carried out at a specialized tertiary care center. Patients with a known history of recent or active infections or inflammatory conditions or missing data were excluded. Baseline clinical and demographic data, laboratory data including blood cell counts, Serum Homocysteine and Folate levels measured before the initiation of anticoagulation, and modified Rankin Scale (mRS) scores at discharge were collected. To study short‐term functional outcomes, patients were dichotomized into good (mRS 2 or less, encoded as output class 1) and poor (mRS >2 encoded as output class 0) functional outcome groups. Machine learning models were built in order to predict if patients belonged to good or poor function outcome groups (mRS2). The classifiers were validated using K‐fold cross validation to pick the best hyperparameters and were tested using a holdout data. The classifiers built include Support Vector Machine (SVM), Logistic Regression Regression (LR), Random Forests (RF) and Gradient Boosting. The performance of the classifiers are compared based on the Area Under the Receiver Operating Characteristic (AUROC) and Sensitivity. Results 162 patients were included in the final dataset, 32 of whom had a significant disability (mRS>2) at discharge. The top variables for predicting the functional outcome groups were Platelet count, Serum Homocysteine, Neutrophil‐to‐Lymphocyte ratio, and Total Leucocyte Count at the time of admission. Support Vector Machine (SVM) was found to be the best‐performing algorithm with an accuracy of 0.86 and AUC of 0.643. Conclusion Using the SVM machine learning algorithm, a predictive model consisting of baseline platelet count, serum homocysteine, total leucocyte count, and the inflammatory marker Neutrophil‐to‐Lymphocyte Ratio was developed to predict short‐term functional outcomes in acute CVT. The high value of accuracy (0.86) indicates that the algorithm is able to reliably detect short‐term functional outcomes.