Modelos de predicción de riesgo de infecciones pulmonares después de un accidente cerebrovascular: una revisión...

Stroke is the leading cause of disability and second leading cause of death worldwide. Pulmonay infection represents the most fatal complication following stoke, accounting for substantial mortality and economic burden. Although various prediction models have been developed to identify susceptible patients, systematic evaluation of their clinical utility to guide therapeutic management is required. In order to use the models that forecast risks for lung infections among patients with stroke as a reference for therapeutic management, this study aims to systematically evaluate them.

Relevant materials were found by searching databases such as the Cochrane Library, PubMed, MEDLINE, Embase, Wanfang, VIP, CBM, and China National Knowledge Infrastructure. Studies on risk prediction models for poststroke pulmonary infections that were published from database's creation to December 2023 were included in the search. The included studies were analyzed and compared in terms of characteristics, research types, predictive factors, model construction methods, and results.

Twenty-four models from 16 studies were included, and the included models' area under the curve values from 0.740 to 0.960, and there were 8 to 32 possible predictors. Age, National Institutes of Health Stroke Scale score, dysphagia, diabetes history, and disturbance of consciousness (Glasgow Coma Scale score) were the most frequently reported predictors. Because of the unreasonably large sample size, the single-factor analysis used to screen predictive features, and the absence of a model performance evaluation, the models demonstrated had not showed a high risk of bias.

Predictive algorithms for the risk of lung infections in patients with stroke are still being studied. Prospective studies should focus on improving research design and reporting, and conducting internal and external validation to develop localized, high-performance, and user-friendly predictive models.