Artificial intelligence outperforms early warning scores at detecting sepsis: a retrospective Danish study

Publikation: KonferencebidragKonferenceabstrakt til konferenceForskningpeer review

Sepsis is a life-threatening condition, and it is essential that the healthcare system quickly identifies patients and treats them adequately. Unfortunately, the early detection of sepsis remains a challenging problem, and even experienced physicians have difficulties in detecting sepsis early and accurately. We aimed to develop an Artificial Intelligence-based Early Warning Score System (AI-EWS) for the early detection of sepsis that is better than the currently used Modified Early Warning Scores (MEWS) and Sequential Organ Failure Assessment (SOFA).

In this register study, we included health data from the years 2010 to 2017 on all citizens 18 years or older with residency in one of four Danish municipalities (Odder, Hedensted, Skanderborg, and Horsens). All relevant hospital contacts from multiple hospitals within the region were identified by the unique national social security system number (1,002,450 contacts). 754,179 non-acute outpatient contacts and 89,202 inpatient contacts with a duration of fewer than six hours were removed. From that set, 134,983 contacts with no episodes of suspected infection were removed leaving 24.076 inpatient contacts included for analysis. After inclusion each inpatient contact underwent a binary classification process to denote them as either sepsis-positive or sepsis-negative. The classification was made based on patients meeting the gold standard for sepsis based on the Third International Consensus Definitions for Sepsis (Sepsis-3). 1,635 (6.8%) inpatient contacts were classified sepsis positive. We included data about biochemistry (blood gas analysis), vital signs, Glasgow Coma Scores, and early warning scores from the electronic health record.
We developed the AI-EWS early sepsis detection model as a deep neural network composed of an embedding layer followed by a temporal convolutional network (TCN). The TCN has four temporal blocks, each with 540 filters of kernel size 10. The dilation rate of the convolutional filters was exponentially increased for each of the stacked temporal blocks. The AI-EWS was trained using Adam optimization, with a learning rate of 0.0005 and a batch size of 200.
The AI-EWS was validated using 5-fold cross-validation. As comparative measures, we used the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC). The model was compared with TOKS (Tidlig Opsporing af Kritisk Sygdom), a Danish MEWS variant, and SOFA.

The following results are reported three, six, and twelve hours before sepsis with mean values and 95% confidence intervals. AI-EWS: (AUROC: 0.88(0.85;0.91), 0.83(0.79;0.87), 0.82(0.79;0.87); AUPRC: 0.41(0.40;0.43), 0.37(0.34-0.39), 0.30(0.26;0.33)), SOFA: (AUROC: 0.77(0.74;0.79), 0.73(0.71;0.74), 0.70(0.65;0-75); AUPRC 0.18(0.16;0.19), 0.16(0.15;0.18), 0.13(0.11;0.16)), and TOKS: (AUROC: 0.68(0.67;0.70), 0.59(0.58;0.59), 0.57(0.55;0.58); AUPRC: 0.12(0.10;0.14), 0.09(0.07;0.10), 0.10(0.09;0.10)). Furthermore, the AI-EWS reduced the number of false positives relatively by 84.6% and 79.4% compared to TOKS and SOFA, respectively, at the same sensitivity of 0.4.

Discussion and conclusions:
The AI-EWS outperformed the SOFA and TOKS in the early detection of sepsis, with an increase in AUROC by 29.4% and AUPRC by 241.7% when compared to TOKS, the currently used early warning tool in Denmark. We conclude that the AI–EWS could be used to improve clinical utility by enabling earlier sepsis interventions and should be tested in a prospective randomized trial.

Trial Registration:
Not registered. Register study.

This work was supported by the Innovation Fond Denmark (case number 8053-00076B).

Ethical approval and informed consent:
Not needed
Udgivelsesår14 okt. 2019
StatusUdgivet - 14 okt. 2019
BegivenhedEUSEM 2019: The European Emergency Medicine Congress - Prague Congress Centre, Prague, Tjekkiet
Varighed: 12 okt. 201916 okt. 2019
Konferencens nummer: 13


KonferenceEUSEM 2019
LokationPrague Congress Centre


  • Sepsis, Machine Learning, Artificial intelligence, Maskinlæring, Kritisk sygdom, Akutmedicin, Forudsigelse, Prædiktion

Se relationer på Aarhus Universitet Citationsformater


  • EUSEM 2019

    Aktivitet: Deltagelse i eller arrangement af en begivenhed - typerDeltagelse i eller organisering af konference


ID: 168634267