Comparison of the Automated Pediatric Logistic Organ Dysfunction-2 Versus Manual Pediatric Logistic Organ Dysfunction-2 Score for Critically Ill Children*

Abstract

Objectives: The Pediatric Logistic Organ Dysfunction-2 is a validated score that quantifies organ dysfunction severity and requires complex data collection that is time-consuming and subject to errors. We hypothesized that a computer algorithm that automatically collects and calculates the Pediatric Logistic Organ Dysfunction-2 (aPELOD-2) score would be valid, fast and at least as accurate as a manual approach (mPELOD-2). Design: Retrospective cohort study. Setting: Single center tertiary medical and surgical pediatric critical care unit (Sainte-Justine Hospital, Montreal, Canada). Patients: Critically ill children participating in four clinical studies between January 2013 and August 2018, a period during which mPELOD-2 data were manually collected. Interventions: None. Measurements and Main Results: The aPELOD-2 was calculated for all consecutive admissions between 2013 and 2018 (n = 5,279) and had a good survival discrimination with an area under the receiver operating characteristic curve of 0.84 (95% CI, 0.81–0.88). We also collected data from four single-center studies in which mPELOD-2 was calculated (n = 796, 57% medical, 43% surgical) and compared these measurements to those of the aPELOD-2. For those patients, median age was 15 months (interquartile range, 3–73 mo), median ICU stay was 5 days (interquartile range, 3–9 d), mortality was 3.9% (n = 28). The intraclass correlation coefficient between mPELOD-2 and aPELOD-2 was 0.75 (95% CI, 0.73–0.77). The Bland-Altman showed a bias of 1.9 (95% CI, 1.7–2) and limits of agreement of –3.1 (95% CI, –3.4 to –2.8) to 6.8 (95% CI, 6.5–7.2). The highest agreement (Cohen’s Kappa) of the Pediatric Logistic Organ Dysfunction-2 components was noted for lactate level (0.88), invasive ventilation (0.86), and creatinine level (0.82) and the lowest for the Glasgow Coma Scale (0.52). The proportion of patients with multiple organ dysfunction syndrome was higher for aPELOD-2 (78%) than mPELOD-2 (72%; p = 0.002). The aPELOD-2 had a better survival discrimination (area under the receiver operating characteristic curve, 0.81; 95% CI, 0.72–0.90) over mPELOD-2 (area under the receiver operating characteristic curve, 0.70; 95% CI, 0.59–0.82; p = 0.01). Conclusions: We successfully created a freely available automatic algorithm to calculate the Pediatric Logistic Organ Dysfunction-2 score that is less labor intensive and has better survival discrimination than the manual calculation. Use of an automated system could greatly facilitate integration of the Pediatric Logistic Organ Dysfunction-2 score at the bedside and within clinical decision support systems.

A Machine Learning Classifier Improves Mortality Prediction Compared With Pediatric Logistic Organ Dysfunction-2 Score: Model Development and Validation

“…There are also many other classification methods and variations of these machine learning models that could be investigated. Of note, PELOD-2 has a lower AUC in our population than in previously studied populations ( 1 , 10 , 11 ). Although we suspect that this difference is attributable to differences in patient populations, and performance of the PELOD-2 logistic regression model improved after retraining using local data, we cannot exclude other systematic biases in the current work.…”

“…PELOD RF , a random forest-based classifier for predicting mortality using the same variables as the PELOD-2 score, had better discrimination and calibration than PELOD-2 for predicting PICU mortality. Although mortality scores are not typically used at the bedside due to complexity of data collection and chance of human error ( 10 ), they are essential for quality assessment and to control for severity of illness in clinical studies ( 1 , 5 , 6 , 10 ). We found that most machine learning models had better performance than the logistic regression-based PELOD-2 model, even after relearning PELOD-2 model on local data.…”

“…The PELOD score was updated to PELOD-2 to include mean arterial pressure (MAP) and lactatemia, which are also variably present in the Sequential Organ Failure Assessment (SOFA) ( 7 ), pediatric SOFA ( 8 ), and Pediatric MODS scores ( 9 ), and hepatic dysfunction was removed ( 1 ). The discrimination (as measured by area under the receiver operating characteristic curve [AUC]) of PELOD-2, which is a multivariable logistic regression model, for predicting mortality ranges from 0.773 to 0.934 in several PICU populations ( 1 , 10 – 12 ). Logistic regression is among the most commonly used prediction methods in biomedicine because it converges on parameter estimates relatively easily, is familiar, and is easy to implement ( 13 ).…”

“…One challenge related to using severity scores such as PELOD-2 is the time required to gather the necessary data ( 17 ). Sauthier et al ( 10 ) created an automatic algorithm to calculate PELOD-2 (called “aPELOD-2” by the authors) from the electronic health record (EHR) that was less labor intensive and had better discrimination for survival than the manually calculated score. This automated method enables researchers to use amounts of data that would be challenging for humans to process in a reasonable amount of time ( 10 ).…”

“…Sauthier et al ( 10 ) created an automatic algorithm to calculate PELOD-2 (called “aPELOD-2” by the authors) from the electronic health record (EHR) that was less labor intensive and had better discrimination for survival than the manually calculated score. This automated method enables researchers to use amounts of data that would be challenging for humans to process in a reasonable amount of time ( 10 ). We aimed to use the larger datasets enabled by aPELOD-2 to develop and validate machine learning models with improved prediction of PICU mortality.…”