SCCM Resource Library-26-Simplified Paediatric Index of Mortality 3 (PIM3) Score by Explainable Machine Learning Algorithm

Video Transcription

Hello everyone, my name is Orkun Baloglu. I am a pediatric intensive care physician at Cleveland Clinic Children's.  I am also a clinical assistant professor at the Lerner College of Medicine Key Western Reserve University. And I work at Cleveland Clinic Clinical Artificial Intelligence.  The titer I talked about today is the name of the Simplified Pediatric Index of Mortality Score with the machine student algorithm that can explain it. I have nothing to explain.  Most of us know that the Pediatric Index of Mortality, or PIM-3, is a tool to combat the risk of mortality in PQ disease. PIM-3 variants  Pupillary Exam Findings, Type of PQ Admission, Mechanical Ventilation in the first 24 hours of PQ Admission, Base Access, Systolic Blood Pressure, FIO-2, PAO-2,  Surgical Recovery Status, and Weighted Diagnostic Category. Shapley Additive Explanations, or SHAP, is a relatively new method for explaining predictions of machine learning models.  A machin student agrees on how beneficial the model is to his latest release. In our experiment, to create a simple version of PIM-3  We create familiar input varios from SHAP analysis. VPS's data for patients who step into Clinic Children's  It has been found for the pawns who step into Clinic Children since 2008 and 2019. Like GBM, this is a gradient-boosting technique used to create machine pupil models.  was used as a model. Red verification has been made up to 10. SHAP varios close to 0.1 or increasing 0.1 varios were used in the machine student model.  Result varios are measures between the 3, 7, 14 and 30 day PQ step and the entire PQ measure. AUC varios were compared to determine the performance of machine student models.  5068 pawns were analyzed. All measurements are close to 1%. The most important input variables of the original PIM-3 variables of the SHAP analysis are the machine student model.  He identified 4 variants that he had experience with. These 4 variolars determined the number of PQ steps, mechanical ventilation in the first 24 hours of the PQ step, the AUC measure, and the machine pupil model.  The final machine student model was compared with the original PIM-3 varios of the machine student model. We can explain these numbers.  As a result, using SHAP analysis, the original PIM-3 variables were reduced to 4 variables and the predictive performance of the machine student model was consequently,  It was delivered to the machine student model with its original PIM-3 pages.  We have our foundations here. Also, my friends, Dr. Eze Tandu, Dr. Latifi and Dr. I thank NASA for participating in this work. Also, for your attention and time  Thank you for listening to this show. You can reach me via my email or Twitter. Thank you very much.

Video Summary

In this video, Dr. Orkun Baloglu presents the Simplified Pediatric Index of Mortality Score (PIM-3) and its application to machine learning models. PIM-3 is a tool used to assess the risk of mortality in pediatric patients. The machine learning model was created using SHAP analysis, which identified four important variables: number of pediatric intensive care (PIC) steps, mechanical ventilation in the first 24 hours of PIC, AUC measure, and the machine pupil model. By reducing the PIM-3 variables to these four variables, the predictive performance of the machine learning model was improved. The final model was compared to the original PIM-3 variables.

Asset Subtitle

Pediatrics, Quality and Patient Safety, 2021

Asset Caption

The Society of Critical Care Medicine's Critical Care Congress features internationally renowned faculty and content sessions highlighting the most up-to-date, evidence-based developments in critical care medicine. This is a presentation from the 2021 Critical Care Congress held virtually from January 31-February 12, 2021.

Orkun Baloglu, MD

Introduction/Hypothesis: PIM3 is a validated tool for assessment of mortality risk in pediatric intensive care unit (PICU). Pupillary exam findings, type of admission, mechanical ventilation in the first 24 hours of admission, base excess, systolic blood pressure (SBP), fraction of inspired oxygen, partial pressure of arterial oxygen, surgical recovery (yes/no), weighted diagnostic category are the input variables of PIM3. SHapley Additive exPlanations (SHAP) is a new method for explaining predictions of machine learning (ML) models by calculating how much each input variable contributes to the final output of a ML model. We aimed to create a simpler version of PIM3 with input variables identified by SHAP analysis.

Methods: Data from the Virtual Pediatric Systems for patients admitted to The Cleveland Clinic Children`s PICU between 2008 and 2019 was obtained. LightGBM (a gradient boosting decision tree algorithm) was used for building the ML models. Stratified 10-fold cross-validation was performed. Input variables with SHAP values â‰¥0.1 were retained and used in the subsequent ML model. Mortality within 3, 7, 14 and 30 days of PICU admission and overall PICU mortality were the output of the predictive models. AUROC values were compared for discriminatory performance of the models.

Results: 5068 patients were analyzed, overall mortality was 1.04%. SHAP analysis of the ML model with the original PIM3 variables revealed 4 variables (type of admission, mechanical ventilation in the first 24 hours of PICU admission, SBP and weighted diagnostic category) as the most important input variables contributing to prediction of the ML model. The final ML model with those 4 variables achieved similar AUROC values compared to the ML model with the original PIM3 variables (0.90 95%CI [0.83-0.98] vs 0.84 [0.68-1.00] p:0.48 for mortality<3d, 0.95 [0.92-0.98] vs 0.93 [0.87-0.98] p:0.46 for <7d, 0.93 [0.90-0.96] vs 0.92 [0.87-0.97] p:0.71 for <14d, 0.91 [0.87-0.94] vs 0.92 [0.90-0.95] p:0.44 for <30 d of PICU admission,0.90 [0.87-0.94] vs 0.91 [0.88-0.94] p:0.87 for overall PICU mortality).

Conclusions: By utilizing SHAP analysis, the original PIM3 variables were reduced to 4 variables and discriminatory performance of the ML model utilizing these 4 variables was non-inferior to the original PIM3 variables for PICU mortality prediction.

Meta Tag

Content Type Presentation

Knowledge Area Pediatrics

Knowledge Area Quality and Patient Safety

Knowledge Level Advanced

Membership Level Select

Tag Scoring Systems

Tag Mortality

Year 2021

Keywords

Simplified Pediatric Index of Mortality Score

machine learning models

SHAP analysis

pediatric intensive care (PIC) steps

mechanical ventilation

	Society of Critical Care Medicine 500 Midway Drive Mount Prospect, IL 60056 USA Phone: +1 847 827-6888 Fax: +1 847 439-7226 Email: support@sccm.org	Contact Us About SCCM Newsroom Advertising & Sponsorship DONATE	MySCCM LearnICU Patients & Families Surviving Sepsis Campaign Critical Care Societies Collaborative	GET OUR NEWSLETTER
© Society of Critical Care Medicine. All rights reserved. \| Privacy Statement \| Terms & Conditions The Society of Critical Care Medicine, SCCM, and Critical Care Congress are registered trademarks of the Society of Critical Care Medicine.