Aims: Community-acquired pneumonia (CAP) causes millions of annual hospitalizations but is a syndromic diagnosis with a diverse array of microbial triggers and patterns of host injury and response. Identification of clinical phenotypes beyond illness severity, the only dimension applied in management guidelines, could advance our understanding of CAP and guide future trials of personalized therapeutic interventions.

Methods: We performed a secondary analysis of a trial enrolling patients presenting to the emergency department from 2016 to 2019. Phenotypes were identified using Kohenen self-organizing maps applied to prespecified demographic and clinical data available by hospital day 1.  This technique is an unsupervised multidimensional clustering technique which accommodates colinear variables, eliminating the need for dimension reduction techniques.

Results: Among 6,848 eligible patients, three phenotypic clusters were identified, correlating with disease severity. Patients with severe pneumonia were younger, had lower P/F ratio, higher blood urea nitrogen, and higher respiratory rate.  ICU length of stay, hospital length of stay, total cost, and 30-day mortality all significantly worsened from cluster 1 (mild pneumonia) to cluster 3 (severe pneumonia).

Conclusions: Current guidelines emphasize the determination of outpatient versus inpatient treatment of pneumonia, the identification of those likely to require ICU admission and critical therapies, and mortality risk.  In this analysis, we validated current clinical severity-based clusters as robust phenotypes utilizing unsupervised machine learning.