The overall objective of this proposed research is the derivation of a biomarker-enhanced artificial intelligence (AI)-based pediatric sepsis screening tool (PSCT) (software) that can be used in combination with the hospital's electronic health record (EHR) system to monitor and assess real-time emergency department (ED) electronic health record (EHR) data towards the enhancement of early pediatric sepsis recognition and the initiation of timely, aggressive personalized sepsis therapy known to improve patient outcomes. It is hypothesized that the screening performance (e.g., positive predictive value) of the envisioned screening tool will be significantly enhanced by the inclusion of a biomarker panel test results (PERSEVERE) that have been shown to be effective in prediction of clinical deterioration in non-critically ill immunocompromised pediatric patients evaluated for infection. It is also hypothesized that enhanced phenotypes can be derived by clustering PERSEVERE biomarkers combined with routinely collected EHR data towards improved personalized medicine.
Background and Rationale Existing automated pediatric sepsis screening tools (PSCT) based on consensus criteria currently used in emergency departments do not improve early recognition and/or inform personalized therapeutic decisions leading to improved outcomes. The Improving Pediatric Sepsis Outcomes (IPSO) initiative found that by including patients that receive treatment, the extended criteria captured not only patients who developed sepsis with organ dysfunction (OD), but also those in whom early sepsis was treated with OD potentially averted. The objective of the proposed effort is to derive and retrospectively validate a biomarker-enhanced AI-based pediatric sepsis screening tool that can be used to screen ED EHR data to improve early recognition, severity stratification, and the timely initiation of personalized sepsis therapy. CTA and its 6 institutional partners jointly propose to establish two de-identified patient registries: 1) the "EHR-data only cohort" (N = 2000) and 2) the "EHR + biomarker data cohort" (N = 400) in support of this objective. Encounter data elements to be abstracted from EHRs for inclusion in these registries include both structured (e.g., time-stamped physiological measurements, treatments, procedures, outcomes) as well as free text notes. Data Analysis and biases All study data, including physiological data extracted from patient EHR and results of biomarker assays will be analyzed using a variety of machine learning algorithms and techniques towards producing a high precision sepsis screening predictive model. Analytic methods involve standard descriptive statistical analysis of predictive classification performance (e.g., AUC, sensitivity/specificity, PPV, etc.).
Study Type
OBSERVATIONAL
Enrollment
12,961
All participating institutions employ either an algorithmic, manual, or combined algorithmic/manual pediatric sepsis screening protocol for patients that present with fever and/or a concern for infection. While the specific parameters tested in screening tools differ, they generally consist of tests for a systemic inflammatory response (e.g. SIRS) and/or organ dysfunction (e.g. SOFA) and/or high susceptibility (e.g. immunocompromised) factors.
Children's National Hospital
Washington D.C., District of Columbia, United States
Effective Expert System-based Pediatric Sepsis Screening Tool (PSCT)
Over a usability test period, by emulation of the logic of experts in a screening tool that cam be continuously improved with experience, achieve a high level of ED workflow usability towards improved early recognition of IPSO sepsis, as perceived by practicing ED clinicians engaged in usability testing.
Time frame: Final 3 months of study period.
High performance Expert System-based Pediatric Sepsis Screening Tool (PSCT)
To derive a high performing (e.g., sensitivity/specificity \> 90%, PPV \> 40%) PSCT to identify patients in the ED meeting IPSO sepsis criteria using early encounter data (e.g. upon receipt of biomarker data within 1st 1-3 hours of presentation).
Time frame: Using "early data" following presentation to ED, e.g., upon receipt of biomarker data within 1st 3 hours of presentation)
Effective sepsis phenotyping for personalized treatment
To show that combined PERSEVERE biomarker and EHR data as clustering features (e.g. using latent class analysis) enhances the detection of clinically useful prognostic phenotypes.
Time frame: Features based on 1st 6 hours following presentation in patients diagnosed with sepsis and treatment protocol initiated.
This platform is for informational purposes only and does not constitute medical advice. Always consult a qualified healthcare professional.