Background: Despite its high lethality, sepsis can be difficult to detect on initial presentation to the emergency department (ED). Machine learning-based tools may provide avenues for earlier detection and lifesaving intervention.

Objective: The study aimed to predict sepsis at the time of ED triage using natural language processing of nursing triage notes and available clinical data.

Methods: We constructed a retrospective cohort of all 1,234,434 consecutive ED encounters in 2015-2021 from 4 separate clinically heterogeneous academically affiliated EDs. After exclusion criteria were applied, the final cohort included 1,059,386 adult ED encounters. The primary outcome criteria for sepsis were presumed severe infection and acute organ dysfunction. After vectorization and dimensional reduction of triage notes and clinical data available at triage, a decision tree-based ensemble (time-of-triage) model was trained to predict sepsis using the training subset (n=950,921). A separate (comprehensive) model was trained using these data and laboratory data, as it became available at 1-hour intervals, after triage. Model performances were evaluated using the test (n=108,465) subset.

Results: Sepsis occurred in 35,318 encounters (incidence 3.45%). For sepsis prediction at the time of patient triage, using the primary definition, the area under the receiver operating characteristic curve (AUC) and macro F₁-score for sepsis were 0.94 and 0.61, respectively. Sensitivity, specificity, and false positive rate were 0.87, 0.85, and 0.15, respectively. The time-of-triage model accurately predicted sepsis in 76% (1635/2150) of sepsis cases where sepsis screening was not initiated at triage and 97.5% (1630/1671) of cases where sepsis screening was initiated at triage. Positive and negative predictive values were 0.18 and 0.99, respectively. For sepsis prediction using laboratory data available each hour after ED arrival, the AUC peaked to 0.97 at 12 hours. Similar results were obtained when stratifying by hospital and when Centers for Disease Control and Prevention hospital toolkit for adult sepsis surveillance criteria were used to define sepsis. Among septic cases, sepsis was predicted in 36.1% (1375/3814), 49.9% (1902/3814), and 68.3% (2604/3814) of encounters, respectively, at 3, 2, and 1 hours prior to the first intravenous antibiotic order or where antibiotics where not ordered within the first 12 hours.

Conclusions: Sepsis can accurately be predicted at ED presentation using nursing triage notes and clinical information available at the time of triage. This indicates that machine learning can facilitate timely and reliable alerting for intervention. Free-text data can improve the performance of predictive modeling at the time of triage and throughout the ED course.

Conflicts of Interest: FB, NWS, SOF, and JDS are vice president of data science, machine learning research scientist, director of nursing, and cofounder and chief medical officer, respectively, at Vital Software, Inc, a company engaged in developing artificial intelligence clinical decision support products for the emergency department.

👁 Figure 1

Figure 1

Receiver operating characteristic curve of…

Figure 1

Receiver operating characteristic curve of sepsis prediction at the time of initial emergency…

👁 Figure 2

Figure 2

Heat maps applied to words…

Figure 2

Heat maps applied to words and subwords of a sample of emergency department…

👁 Figure 3

Figure 3

Sepsis predictive performance of the…

Figure 3

Sepsis predictive performance of the comprehensive model using a test data set, expressed…