Conference Coverage

VIDEO: Software predicts septic shock in hospitalized patients

View on the News

Key next step: Use predictions to improve outcomes

Development of this algorithm is tremendously important and exciting. It is an example of how researchers can use big data to predict patient outcomes and use that information to help deliver better patient care.

The algorithm’s performance so far is laudable and extremely promising, and has great potential to help deliver better care to patients when they need it, but it requires further validation. The potential importance of earlier identification of septic shock is huge.

Dr. Michelle N. Gong professor of medicine and chief of research in the division of critical care at Albert Einstein College of Medicine and Montefiore Medical Center in New York

Dr. Michelle N. Gong

This approach highlights the opportunity we have to leverage the large amount of data that hospitals collect from patients to help identify at-risk patients in a more timely fashion. But having a nice tool is not enough. We need to actually see that clinicians can take this information and use it to improve patient care. Predicting impending septic shock is a trigger, but the key is what is done about the trigger. Can effective interventions be applied to improve patient outcomes? Risk prediction is the first step. The next step is applying appropriate interventions and changing outcomes.

Michelle N. Gong, MD, is professor of medicine and chief of research in the division of critical care at Albert Einstein College of Medicine and Montefiore Medical Center in New York. She had no disclosures. She made these comments in an interview.


AT ATS 2017

– Researchers have devised a program that can predict severe sepsis or septic shock about 12-30 hours in advance of its onset in hospitalized patients, a feat they hope will allow them to apply early interventions to stave off impending sepsis.

“We’d love to see a change in sepsis mortality. Will earlier recognition and implementation of the sepsis bundle – fluids, antibiotics, etc. – improve outcomes?” wondered Heather M. Giannini, MD, in a video interview at an international conference of the American Thoracic Society.

The computer program works by monitoring all the data that enter a patient’s electronic health record during hospitalization. Researchers developed it and designed it specifically for inpatients who are not in the intensive care unit or emergency department.

Results from initial testing during October-December 2015 in 10,448 patients hospitalized at one of three participating Philadelphia hospitals showed the program predicted subsequent severe sepsis or septic shock with a sensitivity of 26% and a specificity of 98%, reported Dr. Giannini, a researcher in the Center for Evidence-Based Practice at the University of Pennsylvania in Philadelphia. Analysis also showed a positive likelihood ratio of 13 for severe sepsis or septic shock actually occurring following an alert generated by the computer program, a level indicating a “very strong” ability to predict sepsis, she said.

Dr. Giannini and her associates developed the prediction program using a technique called “computational machine learning,” an alternative to standard logistic regression modeling that is better suited to analyzing large data sets and can better integrate outlier data points. They took EHR data for all non-ICU, non-ED inpatients at three Philadelphia hospitals during a 3-year period during 2011-2014 and had the program focus particularly on EHR data gleaned from the nearly 1,000 patients who developed severe sepsis or septic shock during the 12 hours preceding the start of these sepsis events. The analysis identified patients as having developed severe sepsis or shock if they had a blood draw positive for infection at the same time as having a blood lactate level above 2.2 mmol/L or a systolic blood pressure below 90 mm Hg.

To create the algorithm the machine-learning device compared the EHR entries for patients who developed severe sepsis or septic shock with EHR data from patients who did not, a process that involved hundred of thousands of data points, Dr. Giannini said. This identified 587 individual types of relevant EHR data entries and ranked them from most important to least important. Important, novel determinants of impending severe sepsis identified this way included anion gap, blood urea nitrogen, and platelet count. The development process also confirmed an important role for many classic markers of septic shock, such as respiration rate, heart rate, and temperature.

The researchers designed the algorithm to have a moderate level of sensitivity to avoid “alert fatigue” from generating too many alarms for impending severe sepsis. Their goal was for clinicians to receive no more than about 10 alerts per day for each hospital.

“We are satisfied with the sensitivity. We felt it was better to have too few alerts rather than overwhelm clinicians. About 10 alerts a day is reasonable,” Dr. Giannini explained. During initial 2015 testing, the system generated a daily average of 11 alerts.

The video associated with this article is no longer available on this site. Please view all of our videos on the MDedge YouTube channel.

On Twitter @mitchelzoler

Recommended Reading

When one patient decompensates, others on the ward may follow
The Hospitalist
Sneak Peak: The Hospital Leader Blog
The Hospitalist
Procalcitonin guidance improves antibiotic stewardship
The Hospitalist
Patient-to-intensivist ratios can influence patient mortality
The Hospitalist
Hospital infections top WHO’s list of priority pathogens
The Hospitalist
Using shock index in the ED to predict hospital admission and inpatient mortality
The Hospitalist
When one patient decompensates, others on the ward may follow
The Hospitalist
Oral iron of no benefit in heart failure with iron deficiency
The Hospitalist
Hospitalists need critical care training pathway
The Hospitalist
HM17 session summary: The hospitalist’s role in the opioid epidemic
The Hospitalist
   Comments ()