Decision Support in the Care of Preterm Newborns-Tool Development

Medical University Of South Carolina, Charleston SC

Investigators

Linked publications & trials

Abstract

DESCRIPTION (provided by applicant): Even after many advances in ventilator management, prediction of extubation outcome for a mechanically ventilated premature infant with respiratory distress syndrome (RDS) remains a challenging task for clinicians. We recently developed a machine-learned model (an Artificial Neural Network, ANN) to assist in decision- making regarding extubation of premature newborns (Mueller et al., 2004, 2006). The ANN model was found to perform with accuracy comparable to that of experienced clinicians;however, this approach needs to be compared to equally powerful machine-learning approaches before it can be evaluated in clinical practice. An appropriately validated decision-support tool could help in reducing the number of days a premature infant spends on a mechanical ventilator, and hence the risk of developing short and long-term side effects of mechanical ventilation, resulting in a corresponding decrease in overall health care costs. In this R21 proposal, we will use several machine-learning approaches combined as a committee formation to obtain the best prediction of extubation success for a given infant. Further, we will build on the previously developed ANN prototype to create an enhanced decision support tool by developing data representation, storage, management, and most important, causal inference, which will enable effective integration of the resulting web-based decision-support tool with clinical practice. This last feature is only possible due to the integrated nature of the proponents themselves, which range from data structure and mathematical modeling experts to experienced neonatologists with a well established working relationship. The proposed effort aims at using advanced modeling tools for translational research by developing a web-based decision-support tool to aid primarily inexperienced clinicians in their decision-making and by promoting interoperability and data exchange among researchers in this field. The critical feature of this infrastructure is its web-based nature, which enables clinicians to evaluate a predictor's accuracy and parametric sensitivity individually for each neonate without having to use any other software than a web-browser. Such a prediction model will be of critical value not only to increase overall clinical accuracy but also to identify effective measures of validity of the original predictions. The overall aim of this study is to develop a high performing web-based prediction system to use as a decision-support tool in clinical practice and to promote interoperability, and thus, data sharing and interaction among researchers in the neonatal community. PUBLIC HEALTH RELEVANCE: Predicting extubation outcome in premature infants on mechanical ventilators remains a challenging task even for experienced clinicians. In the proposed work, we aim to provide a sophisticated web-based tool that uses a machine-learning committee comprised of artificial neural networks (ANN), support vector machines (SVM), naive Bayesian classifiers (NBC), influence diagrams (ID), boosted decision trees (BDT) and multivariable logistic regression (MLR) to assist primarily inexperienced clinicians in the decision-making. For the implementation of this tool we propose to develop an XML schema and RDFS model that can promote interoperability, and thus, data sharing and interaction among researchers in the neonatal community.