Introduction: Prediction of outcome in stroke patients can help both physicians and patients in making treatment decisions and managing prognostic expectations. Machine learning techniques are being increasingly used in the field of medical research. Hypothesis: We hypothesized that models developed with machine learning techniques are useful for predicting long-term functional outcomes in patients with acute ischemic stroke. Methods: The model was developed with a prospective cohort of acute ischemic stroke patients. This cohort registers stroke patients who were admitted within 7 days from the onset of symptoms. For this study, we included all patients admitted between January 1, 2010, and December 31, 2014. We excluded patients with pre-stroke modified Rankin Scale (mRS) score >2 or missing 3-month mRS score. Univariate analysis was performed to guide in variable selection for machine learning models. Machine learning models were trained to classify patients likely to have unfavorable outcome, defined as 3-month mRS score >2. Developed models included artificial neural network, random forest, support vector machine, and logistic regression models. Area under the receiver-operated characteristic curve (AUC) was used to compare effectiveness of each model. Google’s TensorFlow and scikit-learn toolkit were used for training of machine learning models. Results: Total of 3,524 patients were admitted during the study period. After excluding 454 patients with unavailable 3-month mRS scores, 60 patients with pre-stroke mRS score >2, and 87 patients with missing laboratory tests or clinical data, 2,923 patients were finally enrolled. Of the 2,923 patients, 695 (24%) patients had unfavorable outcome (mRS>2) at 3 months. The AUC was 0.888 for Artificial neural network model, 0.810 for random forest model, 0.836 for support vector machine model, and 0.842 for logistic regression model. We calculated the AUC of ASTRAL score for reference, which showed 0.839 in our study group. Conclusion: Machine learning models, particularly the artificial neural network model, achieved high accuracy of prediction for functional outcome in stroke. This study showed the feasibility of machine learning approach for predicting outcomes in stroke.
Read full abstract