Machine-Learning Prediction of Intensive Care Unit Admission Among Adults Presenting With Community-Acquired Pneumonia Using Routinely Collected Clinical Variables: A Systematic Evidence Synthesis and Model-Development Framework

Abstract

Background: Community-acquired pneumonia (CAP) ranges from mild disease manageable outside hospital to rapidly progressive respiratory or circulatory failure requiring intensive care. Accurate identification of patients likely to deteriorate remains difficult. Conventional scores such as the Pneumonia Severity Index (PSI), CURB-65, SMART-COP and Risk of Early Admission to Intensive Care Unit (REA-ICU) provide structured risk assessment, but fixed weighting and limited interaction modelling may restrict performance. Routinely collected electronic clinical data provide an opportunity for machine-learning prediction.
Objective: To synthesize high-quality evidence available through 2021 regarding prediction of intensive care unit (ICU) requirement in adults with CAP and to define a reproducible machine-learning framework using routinely collected clinical variables.
Methods: A structured evidence synthesis was conducted using studies published no later than 31 December 2021 in first-quartile journals relevant to respiratory medicine, infectious diseases, critical care, clinical epidemiology and medical informatics. CAP severity scores, ICU-prediction studies and methodological literature on clinical prediction modelling were examined. Candidate predictors were restricted to variables realistically available at or shortly after emergency-department presentation.
Results: Existing prediction systems consistently identified respiratory rate, blood pressure, oxygenation, mental status, urea, age, comorbidity, radiographic extent and selected laboratory parameters as markers of severe disease. SMART-COP was designed specifically to identify intensive respiratory or vasopressor support and demonstrated high sensitivity in its derivation and validation cohorts. REA-ICU stratified risk of ICU admission within the first three hospital days from approximately 0.7% to 31%. More recent weighted approaches achieved discrimination approaching an AUC of 0.91. Electronic-health-record studies demonstrated that automated multivariable prediction using routinely captured variables was feasible. However, methodological literature showed that machine learning does not inherently outperform logistic regression and that calibration, external validation and clinical utility are essential.
Conclusion: Machine learning offers a plausible method for improving early CAP triage when applied to clinically available variables and evaluated against established severity scores. A robust model should prioritize early availability, prevent information leakage, address missing data within validation folds, compare interpretable and nonlinear algorithms, and report discrimination, calibration and decision-analytic utility. External validation is required before clinical implementation.