Abstract TP104: Functional Prognostication of Patients Receiving Thrombolytics in Acute Ischemic Stroke With Machine Learning

Abstract

Introduction: Accurate outcome prognostication in acute ischemic stroke (AIS) informs essential medical decision-making for clinicians, patients, and families. Neuronal Networks (NN) is a machine learning ML algorithm modeled on the workings of human neurons that has been shown to be useful for prediction tasks in the healthcare setting. We tested whether an NN algorithm could accurately predict long-term neurological outcomes for AIS patients receiving intravenous thrombolysis using common clinical variables. Methods: Patients with AIS treated with intravenous thrombolysis in our health system between July 1, 2020-June 28, 2022 were included in model creation and internal validation. Cases with a modified Rankin scale (mRS) of 0-2 at 90 days (i.e. independence) were classified as “good” outcomes and mRS 3-6 (non-independent) as “poor” outcomes. An NN prediction tool was trained using clinical and interventional variables collected in the first 24 hours of the hospitalization, such as age, gender, and thrombectomy status. Imputation was used for missing values. The final network had a size of 10 layers and a 0.5 decay. The evaluation was made with 10-fold cross-validation. Report followed the TRIPOD statement. Results: We identified 547 AIS patients who received intravenous thrombolysis. Median Age was 70 (IQR: 58-81), and 254 (46.4%) were classified as Female. A total of 115 (20.1%) underwent thrombectomy. Good outcomes were achieved in 354 (64.7%) patients at 90 days. Following the optimization of prediction thresholds through the best informedness values, our final model had a sensitivity of 0.81 and a specificity of 0.78 for detecting good outcomes. The corresponding Area Under the Receiver Operator Curve (AUC-ROC) is 0.85. Conclusion: In conclusion, we developed a prognostic algorithm to predict good outcomes in AIS patients receiving thrombolysis with excellent AUC-ROC performance. External validation of our findings is necessary. NN can potentially improve prognostication of post-stroke functional outcomes, aiding individualized medical decision-making.