PREDICTING INPATIENT LENGTH OF STAY FOR PATIENTS UNDERGOING NEURO-ONCOLOGICAL OPERATIONS: A MACHINE LEARNING PERSPECTIVE FROM A 10-YEAR STUDY AT A NATIONAL TERTIARY NEUROSURGICAL CENTRE

Abstract

Abstract AIMS Length of stay (LOS) following neuro-oncological (NO) operations is essential information for both clinicians, patients and hospital administrators. This study aimed to analyse the LOS of patients following the most common NO surgeries over time and subsequently create a machine learning (ML) algorithm to predict LOS. METHOD Retrospective case series analysis of all NO operations at a tertiary neurosurgical centre from 2012 to 2022. A total of 4700 patients were identified and classified based on 40 operation types. 3 regression ML models were developed to predict the inpatient LOS: XGBoost Regressor, Huber Regressor and Support Vector Machine. Models were trained using 25 stratified KFold cross-validation with an 80:20 train-test split. All models were evaluated via model coverage score, mean absolute error (MAE), mean squared error (MSE) and root mean squared error (RMSE) metrics. All analysis was conducted in R and Python. RESULTS The average patient age was 53.5±15.7 years, with a mean overall LOS of 160.7hrs. The most common operation was an excision of lesion of brain with 2735 operations. The operation with the longest average LOS was extirpation of lesion of meninges of spinal cord at 268.9hrs, with the shortest being biopsy of lesion of brain at 48.1hrs. The best performing model was the XGBoost Regressor with a mean coverage score=0.885 [0.966, 0.803], RMSE=0.784, MAE=0.594, MSE=0.615. CONCLUSIONS ML models can accurately predict LOS and have potential for use in clinical practice to help clinical decision making. Further studies are needed to validate these models on prospective data and additional predictor variables.