Harnessing the power of stream computing
To help improve outcomes, CHOP is building an enterprise-class stream computing cluster, enabling it to ingest, process and analyze thousands of unique data points at high speed.
As a first step, the hospital created a data warehouse to enable it to use historical data to develop and refine its models, and gradually transition from retrospective to real-time analysis.
“We are developing streaming analytics algorithms that can serve as a continuous monitor to identify acute changes in ventilation, such as main-stem intubation,” explains Ahumada. “The goal of developing these algorithms is to use them in clinical decision-support applications to improve safety for our most vulnerable patients.”
He continues: “We collaborated with clinicians and researchers to use historical data to develop a model in MATLAB to detect whether a patient had been intubated optimally. The current model analyzes a number of data points, the most important of which is the amount of carbon dioxide expelled at the end of the respiration cycle. We discovered that subtle variations in end tidal carbon dioxide levels—when combined with factors such as blood pressure—are a strong predictor of whether or not a tube has been optimally placed.”
CHOP feeds sensor readings from its ventilation monitors to a central medical device integration (MDI) layer, which assigns metadata and forwards the information to the stream computing platform.
William Nieczpiel, Development Manager in the Enterprise Analytics and Reporting Team at Children’s Hospital of Philadelphia explains: “The frequency of our readings can vary anywhere from five seconds for end tidal carbon dioxide levels to five minutes for blood pressure measurements, and each device uses different time standards and formats. To account for these differences, assigning a single, accurate master time code to each piece of data is essential. Similarly, we must link raw data from our sensors with the identity of the patient, which we can also achieve in the MDI layer by combining the device serial number with information from our Epic electronic medical records solution.”
By feeding sensor data from 550 beds and thousands of monitors into its stream computing environment, CHOP’s initial pilot model will ultimately be able to identify when a patient has been fitted with a tube—rather than another type of breathing apparatus such as a mask—the type of tube, and if the tube has been positioned optimally.
“Our initial results have achieved 81 percent accuracy with a specificity of 80 percent and a sensitivity of 84 percent to date, and we are confident that this model will become the first of many invaluable tools for decision support,” comments Ahumada. “We co-developed the model with our anesthesia group, and they are already championing internal evaluation of the model as a clinical decision-support tool. Once our stream computing environment is live, we can begin the all-important quality assurance phase of the project.”
He adds: “During our first pilot, we will be comparing the model’s predictions with the results of x-rays and physical examinations for tubes that may not be optimally placed, and using the results to validate the accuracy of the model. Although positioning errors are rare, minimizing their occurrence is a powerful way to improve care outcomes—and we are excited to move the model into production as an additional method of decision support for clinicians.”
Improving care outcomes with analytics
With an enterprise stream computing environment delivering timely, accurate guidance to its clinicians, CHOP is confident that it can deliver even greater improvements to patient care within the tenets of clinical decision support.
“By comparing tiny variations in vital signs that would be extremely difficult for a human to detect, our pilot model is able to predict at a high level of specificity and sensitivity when tube placement is not optimal within the first five minutes of intubation,” says Ahumada.
Nieczpiel adds: “After the solution goes into production, we plan to deliver real-time alerts directly to screens in the patient’s room via our Epic records system, and/or targeted alerts via mobile devices or other modalities. Although our clinicians will ultimately have the final decision about whether or not a tube is placed optimally, we see that our model will be another safeguard to keep our care at outstanding levels.”
In the long term, CHOP’s aim is to create a versatile, enterprise-class stream computing platform that can support a wide variety of use cases.
“We anticipate that our endotracheal tube positioning model will have a very positive impact on patient care, and it’s just the beginning,” explains Ahumada. “The lessons we are learning today will enable us to combine other published research with real-time data to create smarter analytical models. The potential use cases are infinite, but we are already considering ways to operationalize published models for patient deterioration, sepsis prevention, length of stay reduction, and others.”
“Analysis of medical device data will become more widespread and a significant trend in healthcare in the coming years,” says Martin. “With stream computing, we can acquire, analyze and act on our data to improve care outcomes—we’re excited for what the future holds.”
He concludes: “We would like to give special thanks to Ali Jalali, PhD and Jorge Galvez, MD from Children’s Hospital of Philadelphia—Biomedical Informatics in Anesthesia and Critical Care department for their exceptional contributions to the project.”