Good nutrition is central to the management of hospitalized children, and even more critical for the sick infant. It is well recognized that enteral (GI tract) feeds are to be used whenever possible, as this mode of feeding is associated with the best growth and development, and fewer complications such as liver disease and infection that can occur with parenteral nutrition.
However, initiating and advancing enteral nutrition in the hospitalized child is not without its own issues, as perfusion to the GI tract may be compromised, putting these patients at additional risk of necrotic disease. The premature infant and children with cardiac disease are at particular risk. Currently, there is no good method to determine feeding tolerance in this patient population. Practitioners observe for abdominal distension, pain, bloody stools, infection, etc. These findings are late in the course of illness, and may not even be representative of the disease in question. A method to continuously evaluate for enteral feeding tolerance (EFI) would be of great benefit to the care and management of these patients.
The goal of this project is to develop icuFeeds as a clinical decision support system with advanced machine learning algorithms and smart device interfaces that will enable physicians to predict which patients will develop or are developing EFI. We will leverage a large-scale pediatric ICU database contains information collected in more than 5,000 ICU stays every year.
Specific Aim 1: Define feeding intolerance based on retrospective data, including records of interventions such as decreased enteral feeding, oral feeding status, Heme positive stools, and antibiotic use. Analyze EHR and bedside monitoring data to discover features and validate that they have significant population-level correlation to EFI.
Specific Aim 2: Create a pseudo-real-time, cloud-capable database of the features from Aim 1, enabling future smart device applications. Develop machine learning processes within this system to predict whether an individual patient will develop enteral feeding intolerance and/or progressive GI diseases such as NEC. We hypothesize that EHR and bedside monitoring features will be sufficient to train accurate classifiers for predicting EFI.
Specific Aim 3: Create an interface for web-based smart devices capable of monitoring feeding and alerting users to observed and predicted future feeding intolerance using the best classification algorithms from Aim 2.