\u201cThe real goal of everyone who treats children in the ED is to identify that patient, that child … who is going to progress and have sepsis, a life-threatening response to infection, out of all of the children who come with other infectious symptoms,\u201d Alpern said.<\/p>\n
Alpern and colleagues collected data from more than 1.6 million ED visits between Jan. 1, 2016, and Feb. 29, 2020, at five health systems to train a machine learning algorithm to predict patients\u2019 odds of sepsis and septic shock within 48 hours of arriving to the ED. Then they tested it with another 719,298 ED visits.<\/p>\n
Overall, 48.6% of patients were girls, and the median age was 4.7 years (interquartile range, 1.7-10.1 years). The researchers found 0.35% of patients in the training cohort and 0.37% in the testing cohort had sepsis, and 0.15% had septic shock in both cohorts.<\/p>\n
\u201cA big component of the study is that we are utilizing data that is routinely collected in the electronic health record during a child\u2019s initial care in the ED in order to both derive and then validate the study,\u201d Alpern said.<\/p>\n
Alpern and colleagues tested a logistic regression model and a gradient tree boosting model, each with a target sensitivity, or ability to detect sepsis, of 90%.<\/p>\n
The researchers reported robust area under the receiver operating characteristic curve (AUROC) scores for both models, ranging from 0.923 for predicting sepsis and septic shock with logistic regression to 0.936 for predicting sepsis with gradient tree boosting. The gradient tree boosting model produced a positive likelihood ratio of 4.674 and a positive predictive value of 1.7% for predicting sepsis.<\/p>\n
According to Alpern and colleagues, age-adjusted vital signs, emergency severity index and medical complexity were the strongest predictive variables in the analysis.<\/p>\n
Future studies will involve implementation science to determine how to integrate the model with EHRs, Alpern said. She added that the code for the model is accessible online with the study.<\/p>\n
\u201cEven though we were able to derive and validate predictive models that were very robust, we are still going to need additional work before it can be implemented,\u201d Alpern said.<\/p>\n
For more information:<\/h2>\n Elizabeth R. Alpern, MD, MSCE,<\/b> can be reached at pediatrics@healio.com.<\/p>\n\n
\n
Perspective<\/h2>\n Back to Top
\n
\n
\n <\/div>\n
With over 2 million visits and five hospital systems involved in this study, there were diverse and large data for the algorithm to learn from. AUROC, which is used to assess a model\u2019s performance, was high, showing high accuracy. Using these data along with machine learning can help us, as clinicians, to assess and intervene early for best outcomes. I am excited about this \u201caugmented intelligence,\u201d which can help us get to better patient care, early in the course of disease.<\/default:p><\/p>\n\n
Hansa Bhargava, MD<\/strong><\/p>\n\nChief Clinical Strategy and Innovation Officer, Healio<\/li>\n<\/ul><\/div>\n
Disclosures: <\/strong> Bhargava reports no relevant financial disclosures.
\n <\/p>\n<\/p><\/div>\n<\/p><\/div>\n<\/p>\n
\n
\n
<\/p>\n
Ask a clinical question<\/strong> and tap into Healio AI’s knowledge<\/strong> base.<\/p>\n
<\/p>\nPubMed, enrolling\/recruiting trials, guidelines<\/li>\n
<\/p>\n
Clinical Guidance, Healio CME, FDA news<\/li>\n
<\/p>\n
Healio’s exclusive daily news coverage of clinical data<\/li>\n
\n <\/ul>\n
Learn more<\/button>\n <\/div>\n<\/div>\n