Machine learning in risk stratification of prolonged hospital stay after congenital heart surgery

Objectives Extended postoperative hospital length of stay after paediatric cardiac surgery places strain on hospital resources and creates emotional and financial stress for families. However, awareness of which patients will spend relatively longer in hospital remains limited. We propose two machine learning models to predict prolonged length of stay using only electronic health record data before the surgery has taken place. Methods The first model used a fixed definition of prolonged length of stay as exceeding the 75 th percentile of the entire population (17 days). The second applied diagnosis-adjusted threshold, ranging from 8 and 49 days depending on diagnosis complexity. Both models were built using XGBoost and evaluated using the area under the receiver operating characteristic (ROCAUC) . An additional risk stratification was developed by applying probability thresholds on the model output to define four groups based on predicted risk. Results High ROCAUC was achieved by both the fixed-threshold model (88%) and the diagnosis-adjusted model (84%). The four-group risk stratification aligned with test-set outcomes, with prolonged length of stay cases predominantly occurring in the high-risk group and rarely in the low-risk group. Conclusions It is possible to predict prolonged length of stay, either defined using a fixed threshold or a diagnosis adjusted threshold, preoperatively, with only electronic health record data as input. It is moreover possible to stratify the patients who are at highest risk of having prolonged length of stay. Preoperative knowledge of which patients are at risk of prolonged length of stay may facilitate surgical planning and improve patient and parent communication.