Machine learning-based prediction of deviations from DRG-based average length of hospital stay (LOHS) in a Swiss hospital

Background: Since 2012, Switzerland has a prospective hospital payment method based on diagnosis‐related groups (DRGs), reimbursing a flat rate per case. Deviations from the DRG defined length of hospital stay (LOHS) may indicate deficiencies in quality or management and may result in an economic burden for a hospital. Aim: The aim was to identify predictor variables for deviations from DRG average LOHS using machine learning. Methods: The research was based on existing data of a Swiss cantonal Hospital (KSBL). All patients hospitalized for more than one day from 2015 were included in the study. As possible predictors, we included sociodemographic, diseases and treatment-related characteristics. We used penalized linear regression with Ridge and Lasso regularization and ten-fold cross-validation to train each model. Results: A total of 87,706 patients and 114 features were entered into the models. The predictive efficiency of the optimal combinations of features by Lasso and Ridge were almost equal explaining 24% of the total variance in LOHS deviations. Patients who needed geriatric rehabilitation or presented with cerebrovascular disorders stayed longer than expected (β=2.45 and β=0.69 SD above the DRG average). The analysis of the diseases related to the respiratory system (ICD Chapter 10) showed that patients with acute bronchitis, Asthma bronchiale and Influenza had a higher LOHS than expected by DRG. Relevance: Evidence-based hospital management combined with machine learning will enable new strategies to reduce costs and increase the quality empowering personalized health care services.