Purpose Machine learning has recently been shown to increase the predictive power for the outcomes of hemorrhage and patient modified Rankin Scale(mRS) from pre-intervention data. Machine learning techniques have not been assessed for prediction of length of stay. Pre-procedural and post-procedural variables were also assessed to determine predictors of length of stay.

Methods 600 consecutive patients evaluated with CT perfusion for stroke were retrospectively analyzed. Patients were included if thrombectomy was attempted and excluded if the pre-intervention dataset was incompletely documented, 158 patients were included in the study. We compiled a data set using previous randomized controlled trials and retrospective analysis to guide items selected for data collection. Using data available prior to intervention the data set was divided into training and testing sets via sampling with repletion with a sampling ratio of 0.9. Machine learning models including artificial neural network (ANN), support vector machine (SVM), and multivariate linear regression models were created to predict patient length of stay from pre-intervention data. T-test with Sidak multiple testing correction was used to compare models based on root mean squared error generated from model application to unseen data. To assess statistically significantly variables for length of stay, univariate analysis was performed.

Results Length of of stay in this sample had a mean of 9.8±7.9 days. Median LOS was 8 days.In this patient population factors significantly associated with increased LOS were initial NIHSS, ASPECT score, hemorrhage, and infarct size, all with p<0.05.Support vector machine models (RMSE 8.155±1.325) significantly outperformed linear regression, (8.764±1.094) with p=0.001 for RMSE, at prediction of length of stay. Neural network models on average performed significantly worse than either other model (RMSE 15.583±2.457).

Conclusions Machine learning methods outperform multivariate logistic regression at the prediction of length of stay. Predictive analysis for patient length of stay could help hospital utilization and allow for more aggressive measures to prevent hospital acquired conditions.

Disclosures S. Arndt: None. G. Bennett: None. K. Wojcik: None. A. Albar: None. M. Alhasan: None. J. Ma: None. P. Gulotta: None. J. Milburn: None.