ABSTRACT
BACKGROUND: Post-stroke pneumonia is a frequent complication of stroke and is associated with high mortality. Investigators have described its associations with beta-blocker. However, there has been no evaluation of the role of recombinant tissue plasminogen activator (RTPA). We postulate that RTPA may modify the effect of stroke on pneumonia by reducing stroke disability. We explore this using data from neuroprotection trials in Virtual International Stroke Trials Archive (VISTA)-Acute. METHOD: We evaluated the impact of RTPA and other medications in random forest model. Random forest is a type of supervised ensemble tree-based machine learning method. We used the standard approach for performing random forest and partitioned the data into training (70%) and validation (30%) sets. This action enabled to the model developed on training data to be evaluated in the validation data. We borrowed idea from Coalition Game Theory on fair distribution of marginal profit (Shapley value) to determine proportional contribution of a covariate to the model. Consistent with other analysis using the VISTA-Acute data, the diagnosis of post-stroke pneumonia was based on reports of serious adverse events. RESULTS: The overall frequency of pneumonia was 10.9% (614/5652). It was present in 11.5% of the RTPA (270/2358) and 10.4% (344/3295) of the no RTPA groups. There was significant (p<0.05) imbalance in covariates (age, baseline National Institutes of Health Stroke Scale (NIHSS), diabetes, and sex). The AUC for training data was 0.70 (95% CI 0.65-0.76), validation data was 0.67 (95% CI 0.62-0.73). The Shapley value shows that baseline NIHSS (≥10) and age (≥80) made the largest contribution to the model of pneumonia while absence of benzodiazepine may protect against pneumonia. RTPA and beta-blocker had very low effect on frequency of pneumonia. CONCLUSION: In this cohort pneumonia was strongly associated with stroke severity and age whereas RTPA had a much lower effect. An intriguing finding is a possible association between benzodiazepine and pneumonia but this requires further evaluation.