Hospitals are increasingly turning to data science to reduce adverse clinical outcomes, such as decreasing the prevalence of sepsis, the No. 1 cause of unplanned hospital readmissions.
In a webinar sponsored by Intermedix and presented by Becker's Hospital Review, Jonathan Niloff, MD, and Damian Mingle, chief data scientist at Intermedix, discussed how applying advanced analytics to patient demographic, social determinant and community public health data can help hospitals identify risk and prevent sepsis earlier in the care continuum — before a patient enters an intensive care unit.
Intermedix aimed to incorporate data science into an alert system designed to detect adverse clinical events. The alert system is powered by machine learning, a type of artificial intelligence in which a computer learns over time, rather than having to be programmed like typical software. Software developers train a machine learning model by feeding it relevant information from diverse data sources.
Mr. Mingle noted companies like Amazon, Google and Facebook use machine learning to recommend products and posts, based on a constantly changing outlook of user behavior. "What's beautiful, and fascinating, about this is they're doing it on a monster scale. No analyst could knock that out," Mr. Mingle said. "In my judgment, healthcare can, and should, be leveraging the same sort of approach."
The sepsis problem
There are 1.5 million cases of sepsis each year, and this growing medical problem has clinical and financial implications for hospitals, according to Dr. Niloff.
Sepsis is the leading cause of hospitalization and mortality in U.S. hospitals, resulting in 258,000 deaths each year, he said. The complication is also the No. 1 cause of unplanned readmissions, surpassing conditions like heart attack, heart failure, chronic obstructive pulmonary disease and pneumonia, research in JAMA found.
These hospitalizations add up — sepsis accounts for $24 billion in hospitalization costs each year. "Sepsis is a really big problem," Dr. Niloff explained. "But there is now very good evidence that if one can diagnose sepsis earlier — or predict patients who are likely to develop full-blown sepsis — and institute earlier treatment, you can drive better outcomes and reduce costs."
A study in The New England Journal of Medicine, for example, determined sepsis patients who received antibiotics within three hours of presenting at a hospital boasted a lower mortality risk. "Moreover, the effect was hour-by-hour," Dr. Niloff said of the study. "Each hour that antibiotics were received earlier resulted in a measurable improvement in outcome."
Yet catching sepsis early is difficult. Sepsis symptoms are notoriously ambiguous and frequently "suggestive" of other acute conditions, according to Dr. Niloff. "Sepsis is a particularly difficult diagnostic challenge," he said, noting by the time a patient is diagnosed with sepsis, it may be too late for early intervention.
How disparate data informs sepsis prediction
The team at Intermedix chose to apply its existing interest in machine learning to early sepsis intervention.
To develop the sepsis-focused predictive algorithm, Intermedix gathered relevant research on infection risk factors, which Mr. Mingle referred to as the "known universe" for sepsis. This research comprised more than 90,000 articles, from which the Intermedix team's machine learning model identified 2,200 relevant risk attributes for sepsis.
The team, which has already begun deploying the solution, refines and customizes the predictive model at each facility, integrating the past three years of available historical claims data for every hospital using the software.
The goal is for the model to learn how patients at various hospitals might present for sepsis with slightly different risk factors or symptoms. "That's what the notification's built on," Mr. Mingle explained.
How the predictive algorithm embeds into clinical workflow
Once an emergency department physician recommends a patient be admitted, a copy of the patient's admission, discharge and transfer feed is automatically sent to Intermedix's Condition Awareness solution — without any additional data entry. "We're not asking the clinical team to enter a bunch of [additional] information into the EHR," Mr. Mingle noted. Rather, the solution avoids common pitfalls of an EHR and fits into a clinical workflow hours before that manual data entry would take place.
The Intermedix solution applies the sepsis algorithm to the ADT feed, which evaluates the patient's risk for developing sepsis in the context of the research-based risk attributes and the hospital's diagnostic history. If the solution determines the patient is at high risk for sepsis, it will notify the appropriate clinical staff via secure email or text message within minutes of admission, registration or arrival.
In case studies discussed during the webinar, the solution had a 96.3 percent positive predictive value for early sepsis identification. "You're not going to have your physicians suffering from alert fatigue," Dr. Niloff explained. "When you get one of these alerts, there is a very high probably that the patient is indeed at risk and needs intervention. It's a tool that clinicians will come to respect and find a lot of value from."
Mr. Mingle emphasized the tool is a type of decision support — it supplies information, but does not dictate how clinicians should respond. "We really stay out of the business of telling physicians what we think they ought to be doing," Mr. Mingle said. "We leave the decision of what to do with the information up to the clinical team."
"It's an early notification system," he added. "If we can provide help to a clinical team by an hour — or two, three, four, five — that's tremendous."
Watch the webinar recording here.