Machine learning based prediction of length of stay in acute ischaemic stroke of the anterior circulation in patients treated with thrombectomy

Background: Length of stay is an important factor for managing the limited resources of a hospital. The early, accurate prediction of hospital length of stay leads to the optimized disposition of resources particularly in complex stroke treatment.

Objective: In the present study we evaluated different machine learning techniques in their ability to predict the length of stay of patients with stroke of the anterior circulation who were treated with thrombectomy.

Material and methods: This retrospective study evaluated four algorithms (support vector machine, generalized linear model, K-nearest neighbour and Random Forest) to predict the length of hospitalization of 113 patients with acute stroke who were treated with thrombectomy. Input variables encompassed baseline data at admission, as well as periprocedural and imaging data. Ten-fold cross-validation was used to estimate accuracy. The accuracy of the algorithms was checked with a test dataset. In addition to regression analysis, we performed a binary classification analysis to identify patients that stayed longer than the mean length of stay.

Results: Mean length of stay was 10.7 days (median 10, interquartile range 6-15). The sensitivity of the best-performing Random Forest model was 0.8, the specificity was 0.68 and the area under the curve was 0.73 in the classification analysis. The mean absolute error of the best-performing Random Forest Model was 4.6 days in the test dataset in the regression analysis.

Conclusion: Machine learning has potential use to estimate the length of stay of patients with acute ischaemic stroke that were treated with thrombectomy.