Abstract

Background: With the growing ease of collecting and analyzing large quantities of data, machine learning algorithms have risen in prominence for the purposes of creating models that can discover previously hidden relationships between large amounts of data points, and which can then be used for classification and predictive tasks. Methods: A retrospective database of patients admitted with acute ischemic stroke at a single institution was reviewed. Patients with a modified Rankin Scale (mRS) at discharge and those who had not undergone mechanical thrombectomy were selected. Modified Rankin Scale was dichotomized into good (0 to 2) and bad (3 to 6). 18 total variables, including patient demographics and admission lab values, were selected as inputs. Feature selection was initially performed using boosted decision trees to select the five most salient variables. Patients missing data from these features were dropped. Several classifiers were ensembled together and trained using 10-fold stratified cross validation to determine the likelihood a patient would achieve a good or bad mRS at discharge. Each fold’s accuracy, ROC curve and area under the ROC curve were calculated, and then averaged together once cross validation was complete. Results: 6,320 patients with the diagnosis of acute ischemic stroke from 2006 to 2018 were examined. Of these, only 3,387 had an mRS at discharge recorded, and 1,972 patients were not missing any data and were available for analysis. The most important variables were age, NIHSS on admission, ambulatory status prior to admission, IV tPA administration, and admission INR. The final ensembled model achieved an accuracy of 80.8%, with an AUC 0.88 +/- 0.02. Conclusion: By utilizing machine learning strategies, large sets of data can be analyzed to develop models which can help inform clinicians the likelihood of a patient’s good outcome. Additionally, by examining which features are the most salient to a model, previously un appreciated factors can be highlighted and further studied.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call