Decision tree model development and in silico validation for avoidable hospital readmissions at 30 days in a pediatric population.

Abstract

Background and objective: identifying patients at high risk of avoidable readmission remains a challenge for healthcare professionals. Despite the recent interest in Machine Learning in this topic, studies are scarce and commonly using only black box algorithms. The aim of our study was to develop and validate in silico an interpretable predictive model using a decision tree inference to identify pediatric patients at risk of 30-day potentially avoidable readmissions.

Methods: a retrospective cohort study was conducted with all patients under 18 years admitted to a tertiary university hospital. Demographic, clinical and nutritional data were collected from electronic databases. The outcome was the potentially avoidable 30-day readmissions. The J48 algorithm was used to develop the best-fit trees capable of classifying the outcome efficiently. Leave-one-out cross-validation was applied and we computed the area under the receiver operating curve (AUC).

Results: the most important attributes of the model were C-reactive protein, hemoglobin and sodium levels, besides nutritional monitoring. We obtained an AUC of 0.65 and accuracy of 63.3 % for the full training and leave-one-out cross-validation.

Conclusion: our model allows the identification of 30-day potentially avoidable readmissions through practical indicators facilitating timely interventions by the medical team, and might contribute to reduce this outcome.