Setting up standards: A methodological proposal for pediatric Triage machine learning model construction based on clinical outcomes.

• We deal with triage prediction systems using machine learning in pediatric patients. • A very large in-house dataset was used for model construction and validation. • We propose a data re-labeling process aiming to improve clinical outcome prediction. • Random forest provides best results over a bunch of multi-class performance metrics. • This result outperforms traditional Triage models. Triage is a critical process in hospital emergency departments (ED). Specifically, we consider how to achieve fast and accurate patient Triage in the ED of a pediatric hospital. The goal of this paper is to establish methodological best practices for the application of machine learning (ML) to Triage in pediatric ED, providing a comprehensive comparison of the performance of ML techniques over a large dataset. Our work is among the first attempts in this direction. Following very recent works in the literature, we use the clinical outcome of a case as its label for supervised ML model training, instead of the more uncertain labels provided by experts. The experimental dataset contains the records along 3 years of operation of the hospital ED. It consists of 189,718 patients visits to the hospital. The clinical outcome of 9271 cases (4.98%) wa hospital admission, therefore our dataset is highly class imbalanced. Our reported performance comparison results focus on four ML models: Deep Learning (DL), Random Forest (RF), Naive Bayes (NB) and Support Vector Machines (SVM). Data preprocessing includes class imbalance correction, and case re-labeling. We use different well known metrics to evaluate performance of ML models in three different experimental settings: (a) classification of each case into the standard five Triage urgency levels, (b) discrimination of high versus low case severity according to its clinical outcome, and (c) comparison of the number of patients assigned to each standard Triage urgency level against the Triage rule based expert system currently in use at the hospital. RF achieved greater AUC, accuracy, PPV and specificity than the other models in the dychotomic classification experiments. On the implementation side, our study shows that ML predictive models trained according to clinical outcomes, provide better Triage performance than the current rule based expert system in operation at the hospital. [ABSTRACT FROM AUTHOR]