AI Case Study
Children's Hospital of Los Angeles predicts when to discharge patients from pediatric intensive care using deep recurrent neural networks
Children's Hospital of Los Angeles analysed the history of 5,500 patient vitals from pediatric intensive care unit (ICU) episodes to better predict when to discharge. They build four predictive models from basic regression to recurrent neural networks and found the deep neural networks performed well.
Industry
Healthcare
Healthcare Providers And Services
Project Overview
The team tried four models of age-normal regression, polynomial regression and two neural networks. The data for the supervised learning models was "6,899 Pediatric Intensive Care Unit (PICU) episodes represented by 5,464 patients which were collected between 2009 and 2016."
"Each episode had time series measurements for 375 variables representing vitals, laboratory results, interventions and drugs."
Two deep recurrent neural network (RNN), residual LSTMs "were developed to predict each individual patient’s ICU-PASS values." The first model was "trained on all the time points before medical discharge of each episode in the training set, while the second (RNN12h) trained only on data through the first 12 hours of those same episodes."
Reported Results
"Machine learning approaches to EMR medical data are becoming
more common and have been increasingly reported. We have demonstrated that an advanced deep learning methodology using the rich data in the EMR, rather than simply using a relationship between age and vital signs, more accurately predicted individual patient’s PASS for discharge from the PICU."
Technology
The team tried four models of age-normal regression, polynomial regression and two neural networks. The two deep recurrent neural network (RNN), residual LSTMs "were developed to predict each individual patient’s ICU-PASS values. The models used data from the pre-medical discharge period ...to predict ICU-PASS values. The first model (RNNPMD) was trained on all the time points before medical discharge of each episode in the training set, while the second (RNN12h) trained only on data through the first 12 hours of those same episodes."
Function
R And D
Core Research And Development
Background
Children's Hospital of Los Angeles wanted to better understand whether they could predict patient-specific vitals that would be acceptable for their discharge from a pediatric intensive care unit.
"Heart rate, systolic blood pressure, and diastolic blood pressure were selected as vitals critical to determining physiologic stability"
Benefits
Data
"The data were extracted from anonymized observational
clinical data collected in Electronic Medical Records (EMR,
Cerner) in the Pediatric Intensive Care Unit (PICU) of Children’s
Hospital Los Angeles between 2009 and 2016. "
"6,899 Pediatric Intensive Care Unit (PICU) episodes represented by 5,464 patients collected between 2009 and 2016.
Each episode had time series measurements for 375 variables representing vitals, laboratory results, interventions and drugs
They also included a time indicator for PICU medical discharge and physical discharge."