#1387 Machine learning identifies variables associated with risk of hospitalisation in a national in-centre hemodialysis patient cohort

Abstract Background and Aims Unplanned emergency hospitalisation is associated with a loss of quality of life and reduction in life expectancy for patients receiving hemodialysis treatment. Artificial intelligence tools including machine learning allow inferences to be made from large datasets on the patient characteristics that might be associated with the risk of hospitalisation. Method Machine learning models, including XGBoost, Random Forest, Support Vector Machines (SVM), and LASSO Regression, were developed using 18 variables from a longitudinal database of approximately 30,000 dialysis patients across 32 Davita centres in Colombia over two years. The data was stratified into a 70/30 train-test split and further stratified into four risk groups based on the probability of hospitalisation in the next month: Low, Medium, High, and Critical (0%–50%, 50%–75%, 75%–90% and 90%–100% respectively). Minimal feature engineering was applied to maintain interpretability for clinicians. The best-performing model was validated using a separate dataset collected prospectively over a ten-month period. Results XGBoost demonstrated the highest performance among models, achieving an AUC of 0.72 (0.70–0.74) on the test set in predicting hospitalisation. However, LASSO Regression, with a comparable AUC of 0.71 (0.70–0.72), was selected for validation due to its simplicity and improved feature interpretability. On the validation set, LASSO achieved an AUC of 0.85 (0.77–0.91). While the model struggled to differentiate low-risk patients, it accurately identified critical-risk patients with a sensitivity of 86% and specificity of 40%. Key predictors of hospitalisation included anaemia, missed treatment rate, vascular access type, and albumin levels, with patients in the high and critical risk groups having poorer/less clinically favourable results for these variables. This was reflected in the lower specificity due to the model's tendency to classify patients as high risk based on these variables. Conclusion Machine learning models can help identify variables associated with a high risk of hospitalisation in patients undergoing in-centre hemodialysis. The high importance variables identified in our study are mechanistically related to outcomes and it is important to maintain face validity for multidisciplinary healthcare professionals when scrutinising such models. It is currently being assessed whether such models can actively reduce hospitalisation rates through proactively managing variables associated with hospitalisation risk. Further work is required in addressing issues such as bias and replicability of findings.