Hospital leans on machine learning to reduce sepsis-related mortality rate

Cabell Huntington Hospital also diminished the average sepsis-related hospital length of stay with machine learning-generated clinician alerts.
By Bill Siwicki
03:57 PM
Share
machine learning reduces sepsis mortality rate

Cabell Huntington Hospital in Huntington, West Virginia. Credit: Google Maps

Last year, Cabell Huntington Hospital faced sepsis head-on and came out on top. Implementing machine learning technology specifically designed to fight sepsis in part through clinician alerts, the organization saw the sepsis-related in-hospital mortality rate was 33.5 percent lower during the post-implementation period and the average sepsis-related hospital length of stay was 17.1 percent lower during the same period. Analyses included 2,298 adult patients in the emergency department and intensive care unit.

Through an ongoing review of internal data, it appears that InSight clinical alerts, from machine learning vendor Dascena, and clinical documentation/coding of sepsis are showing an increased correlation, said Hoyt J. Burdick, MD, chief medical officer at Cabell Huntington Hospital.

“Of course, this phenomenon is not just dependent on the machine logic alerting, but is also subject to clinician education, documentation, coding and billing variables,” he explained. “But since we only recently began to adjust some of the machine logic parameters, it seems more likely that the clinicians are more confident in making diagnoses and decisions based upon the improved alerts.”

[Also: Lafayette General earns HIMSS Davies Award for slashing sepsis rates with tech tools]

A less frequent, more accurate alert via a telephone call may have more impact than an EHR alert that is often false, Burdick added. “We continue to retrospectively review patient charts where InSight alerts fire to monitor sensitivity and specificity compared to clinician diagnosis and documentation,” he said.

Cabell Huntington Hospital clinicians and I.T. staff initially built an electronic Systemic Inflammatory Response Syndrome (SIRS) and sepsis alerting system in Cerner based upon a rules-based model they learned about from Methodist Memphis Le Bonheur.

There were two types of alerts: 1) A SIRS alert that appears in the EHR when a patient has three or more critical vital signs or certain laboratory result values, and 2) A sepsis alert when a patient meets two or more SIRS criteria and also has a lab result indicating organ dysfunction. These alerts were intended to prompt clinicians to initiate a sepsis treatment bundle.

[Also: How Halifax Health combats sepsis: Surveillance, analytics, communications tech]

However, the rules-based system alerts very frequently had a high false positive rate. Cerner subsequently adopted and made available serial versions of a similar rules-based alerting system called St. John Sepsis Surveillance Agent. Some elements of the St. John alert are customizable, and teamwork with I.T. staff at Cabell was critical in adjusting variables such as turning the alerting off for 24 hours after firing, building the primary nursing notification template and establishing order sets in accordance with the most current version of the Surviving Sepsis Guidelines.

“There were also areas of the hospital, like labor and delivery, where the SIRS/Sepsis alert had to be inactivated because of the frequency of firing for normal obstetrical patients,” Burdick explained. “We continued to experience a little over 1000 electronic sepsis alerts from the SIRS-based algorithms with a diagnosed rate of patients with severe sepsis/septic shock of under 100 per month.”

This led the clinical and I.T. team to investigate alternative EHR-based sepsis alerts with goals of earlier detection, increased true positives (high sensitivity) and decreased false positives (high specificity – higher true negatives). These challenges led Cabell to search for a better system and evaluate alternatives like Dascena’s InSight machine logic software.

InSight collects basic data from the EHR and computes sepsis risk by comparing the patient data profile and trends against statistical patterns from its reference library of millions of prior cases of patients with sepsis. InSight runs in the background, so there is no additional action needed from clinicians.

“The software requires a vital signs feed, but also analyzes certain laboratory results, when available,” Burdick said. “When InSight determines that a patient’s data profile is similar to the reference population with sepsis, the software notifies the charge nurse on duty through an automated call to a dedicated phone line per unit. Following the alert, the charge nurse can then begin the sepsis assessment and inform the physician.”

The machine learning technology contributes to heightened accuracy in sepsis screening and earlier detection, he added. And Cabell still operates the St. John Sepsis Surveillance Agent alerts in Cerner.

“Other systems beyond Dascena we have reviewed are rules-based and are triggered by patients who meet a predetermined set of criteria,” he said. “While rules-based algorithms are good at capturing all patients with abnormal vital signs and/or lab result combinations, the alerts tend to be frequent and nonspecific. This contributes to alert fatigue among clinicians, who then may be less likely to believe or act immediately on a true sepsis alert.”

Machine learning also has allowed for improved predictive power in sepsis detection by warning Cabell clinicians of sepsis trends prior to onset of organ dysfunction. This time factor is important since early antibiotic administration is key to reducing sepsis mortality.

Twitter: @SiwickiHealthIT
Email the writer: bill.siwicki@himssmedia.com