The pneumonia severity index: Assessment and comparison to popular machine learning classifiers

Abstract

Introduction: Pneumonia is the top communicable cause of death worldwide. Accurate prognostication of patient severity with Community Acquired Pneumonia (CAP) allows better patient care and hospital management. The Pneumonia Severity Index (PSI) was developed in 1997 as a tool to guide clinical practice by stratifying the severity of patients with CAP. While the PSI has been evaluated against other clinical stratification tools, it has not been evaluated against multiple classic machine learning classifiers in various metrics over large sample size.

Methods: In this paper, we evaluated and compared the prediction performance of nine classic machine learning classifiers with PSI over 34,720 adult (age 18+) patient records collected from 749 hospitals from 2009 to 2018 in the United States on Receiver Operating Characteristic (ROC) Area Under the Curve (AUC) and Average Precision (Precision-Recall AUC).

Results: Machine learning classifiers, such as Random Forest, provided a statistically highly(p < 0.001) significant improvement (∼33% in PR AUC and ∼6% in ROC AUC) compared to PSI and required only 7 input values (compared to 20 parameters used in PSI).

Discussion: Because of its ease of use, PSI remains a very strong clinical decision tool, but machine learning classifiers can provide better prediction accuracy performance. Comparing prediction performance across multiple metrics such as PR AUC, instead of ROC AUC alone can provide additional insight.