Study Design and Setting
A retrospective analysis was performed at Vanderbilt University Medical Centers (VUMC) in Nashville, TN, USA and the Henry Ford Hospital System (HFHS) in Detroit, MI, USA. VUMC is a level 1 trauma center with approximately 125,000 annual ED visits and HFHS includes 9 Eds, 4 of which are free-standing, and 5 hospitals that experience approximately 460,000 annual ED encounters.
Selection of Participants
Data from VUMC were used for the derivation cohort (January 2013 to September 2017) and data from HFHS was used for the validation cohort (January 2014 to January 2018). Eligible patients were selected in similar manner from VUMC’s Synthetic Derivative, a de-identified mirror-image of the system’s electronic health records, by Nashville Biosciences, a subsidiary of VUMC established to support translational research, and from HFHS’ electronic health records (EPIC, Verona, WI). Eligible patients were 18 years of age or older, had a diagnosis of acute pancreatitis (International Classification of Diseases [ICD], 9th Revision, code 577.0 or ICD, 10th Revision, code K85), and required hospital admission within 24 hours of presentation in the ED. Only the first encounter with a patient over the study assessment period was included to ensure no patient was represented more than once in the data set. We excluded patients with chronic pancreatitis, hepatic failure, or with a serum lipase <3x upper limit of the normal reference range. While prognostic assessment of patients discharged from the ED may be useful, we did not include these patients given high rates of concomitant chronic pancreatitis and reduced opportunity to obtain outcomes data given the retrospective nature of the study.
Data Collection and Processing
Data collected for the derivation and validation cohorts included subject-level information on demographics; relevant comorbidities based on ICD codes; diagnoses of respiratory failure, sepsis, acute kidney injury (AKI), and other relevant sequelae as per ICD codes; mechanical ventilation determined by procedure codes for invasive mechanical ventilation; hospital procedure and visit records; mortality; and the first ED-recorded lab values and vital signs. Mechanical ventilation did not include non-Invasive modes of ventilation.
Data analysts pulled electronic health data using standardized methods within each health system. For the validation cohort only, additional chart abstraction was performed to assess the accuracy of electronic data collection for the diagnosis of acute pancreatitis. Physician chart abstractors were blinded to the study hypothesis. All abstractors were trained before initiating chart abstraction and were provided with a data collection manual that including variable definitions and details. Within the validation cohort, physicians abstracted 10% of study charts to authenticate electronic data capture and to test agreement with electronic chart abstraction for the diagnosis of acute pancreatitis. The K coefficient for the diagnosis of acute pancreatitis was 0.94. For missing data, it was assumed that a diagnosis or death was not present if not documented.
Outcome Measures
The primary outcome was 30-day mortality. Secondary outcomes included 180-day mortality; intensive care unit (ICU) admission; length of hospital stay; presence of sepsis, respiratory failure, or AKI; and need for dialysis or mechanical ventilation. To determine these latter outcomes, we used ICD codes documented during the patient’s index hospitalization for acute pancreatitis.
Primary Data Analysis
We selected a priori predictor variables known to be associated with mortality from acute pancreatitis based on existing disease models.8,10,14,15 The variable selection was refined based on their availability during routine ED evaluation and consistent documentation within electronic health records. The a priori predictor variables included age >60 years, gender, self-reported race/ethnicity, the Charlson comorbidity index,16 the presence of ≥2 SIRS criteria (heart rate >90/min, respiratory rate >20/min, temperature >38.0°C or <36.0°C, white blood cells >12x109/L, or white blood cells <4x109/L), the presence of SpO2 saturation <96%, and hematocrit >44%. Of importance, the SpO2 cut-off of <96% was used rather than <95% as it includes the descending portion of the oxygen saturation curve.
For the derivation cohort, these variables were analyzed for their association with 30-day mortality with univariate analysis and retained variables for the final model with p<0.10. A complete case analysis was also performed. To reduce bias in parameter estimates given the low event rate, multivariable logistic regression with Firth’s penalized likelihood estimate was utilized.17 To determine the most parsimonious predictor model, the Akaike Information Criterion was assessed. In addition, model calibration using graphical assessment by loess smoothers18 and model discrimination with area under the receiver operating characteristic curve (c statistic) were evaluated.
The final three variables included in the model were SpO2 <96%, age >60 years, and ≥2 SIRS criteria (Table 1). Each variable was equally weighed and contributed 0 or 1 points to an overall score based on the presence or absence of these three variables resulting in an ED-SAS range of 0 to 3 points. We named the score ED-SAS for SpO2, age, and SIRS. Logistic regression for these score categories with 0 as the reference to assess the odds of death and all secondary outcomes within the derivation data set was performed and applied this derived predictor score to the validation sample to assess its association with primary and secondary outcomes. No model recalibration was performed within the validation data. All analyses were performed with SAS 9.4 (Cary, NC).
Table 1
Parameter
|
Score
|
SpO2 <96%
|
1
|
Age >60 years
|
1
|
≥2 SIRS criteria*
|
1
|
Total Score, range
|
0 – 3
|
*Presence of 2 or more of the following: fever >38.0°C or hypothermia <36.0°C, tachycardia >90 beats/minute, tachypnea >20 breaths/minute, Leukocytosis >12*109/L, or leucopoenia <4*109/L.
ED-SAS, Emergency Department SpO2, Age, and SIRS; SIRS, systemic inflammatory response syndrome
|