Self-organising map clustering identifies high-risk clusters of post-acute mortality in a prospective multicentre study of community-acquired pneumonia

Background Community-acquired pneumonia (CAP) is a leading cause of morbidity and mortality. While tools predicting short-term prognosis exist, there is urgent need for the early identification of patients requiring close follow-up monitoring for post-acute mortality. We therefore conducted cluster analysis of baseline clinical data to investigate predictors of post-acute mortality in CAP. Methods We analysed 7840 participants from the German CAPNETZ cohort, using Self Organising Map (SOM)-clustering and survival analyses. Random survival forests (RSF) were used to identify key predictors of mortality, which then were analysed using time-dependent AUC and Cox proportional hazard regression models. Results SOM-clustering based on ten predictors identified 879 (12%, in 4 clusters) patients with high risk for post-acute (30–180d) mortality. Across the cohort, age and urea were the most important predictors of post-acute mortality, while, in the high-risk cohort, BMI emerged as the strongest predictor, as identified by RSF-modelling. In one high-risk cluster, there was an association with elevated platelet counts (HR: 1.13, 95% CI. 1.03–1.21, p=0.01; increments of 40 platelets/nl, c14, 35% of high-risk patients), in another (c15, 50% of high-risk patients) with elevated urea (HR: 1.06, 95% CI. 1.01–1.11, p=0.02) and CRP (1.27, 95% CI 1.01–1.58, p=0.04. Conclusion Using ten clinical predictors for post-acute mortality in CAP, predictive SOM- clustering revealed several high-risk subgroups, with heterogeneous biomarkers, suggestive of differences in the underlying pathophysiology (thrombocytes, urea, CRP). Adapting medical therapy to these high-risk subgroups may reduce post-acute mortality following CAP.