Representation learning for the prediction of length of stay in infants with congenital heart diseases undergoing cardiac surgery

Abstract Aims Congenital heart disease is the most common birth defect and significantly impacts health, mortality, and healthcare costs. Accurately estimating intensive care unit (ICU) and hospital stays after cardiac surgery could facilitate family planning and optimal healthcare resource allocation. Although some perioperative risk factors have been associated with prolonged stays, tools to estimate the length of stay for the target population remain insufficient. This study employs explainable representation learning to enhance length of stay prediction through comprehensive spatial distribution and local learning. Methods and results Demographic and perioperative data from infants undergoing cardiac surgery were collected. Unsupervised multiple kernel learning was used to reduce data dimensionality and position patients by similarity. K-Means clustering was applied to identify phenogroups. Within each phenogroup, a binary classifier predicted prolonged stays, and a regressor estimated stay durations in days. Algorithms were validated using an independent cohort and compared with current clinical practices. Among 268 episodes of care (45.2% female, median age 0.6 years), 21% had prolonged ICU stay, and/or prolonged hospital stay. Key variables surgery duration, or extra corporeal circulation time were significantly correlated with latent dimensions. Classifiers reaches an AUC of 0.71 in identifying prolonged stays in the ICU and in the hospital. Regressor estimated both length of stay with a median absolute error of one day. This approach outperforms clinicians in both, classification and regression tasks. Conclusion Explainable representation learning leveraging patient similarity provides an interpretable tool for estimating the length of stay in the target population.