The predictive effect of different machine learning algorithms for pressure injuries in hospitalized patients: A network meta-analyses

Pressure injury has always been a focus and difficulty of nursing. With the development of nursing informatization, a large amount of structured and unstructured data has been generated, and it is difficult for traditional methods to utilize these data. With the intersection of artificial intelligence and nursing, it has become a new trend to apply machine learning algorithms to build pressure injury prediction models to manage pressure injuries. However, there is no evidence on the effectiveness of the method and which of a large number of algorithms for machine learning is more applicable to pressure injuries. This review aims to systematically synthesize existing evidence to determine the effectiveness of applying machine learning algorithms for pressure injury management, to further evaluate and compare pressure injury prediction models constructed by numerous machine learning algorithms, and to derive evidence for the best algorithms for predicting and managing pressure injuries. Systematic review and network meta-analysis. A systematic electronic search was conducted in the EBSCO, Embase, PubMed, and Web of Science databases. We included all retrospective diagnostic accuracy trials and prospective diagnostic accuracy trials constructing a predictive model by machine learning for pressure injuries up to December 2021. Two review authors independently selected relevant studies and extracted data using the Cochrane handbook for systematic reviews of diagnostic test accuracy. The network meta-analysis was conducted using statistical software R and STATA. The certainty of the evidence was rated using the QUADAS-2 tool. Twenty-five clinical diagnostic trials with a total of 237397 participants were identified in this review. The results of our study revealed that pressure injury machine learning models can effectively predict these injuries. Combining the algorithms separately yields the main results: decision trees (sensitivity: 0.66, 95% CI: 0.42 to 0.84, specificity: 0.90, 95% CI: 0.78 to 0.96, diagnostic odds ratio [DOR]: 18, 95% CI: 7 to 49, AUC: 0.88, 95% CI: 0.85 to 0.91), logistic regression (sensitivity: 0.71, 95% CI: 0.60 to 0.80, specificity: 0.83, 95% CI: 0.75 to 0.89, DOR: 12, 95% CI: 9 to 17, AUC: 0.84, 95% CI: 0.81 to 0.87), neural networks (sensitivity: 0.73, 95% CI: 0.55 to 0.86, specificity: 0.78, 95% CI: 0.65 to 0.87, DOR: 9, 95% CI: 5 to 19, AUC: 0.82, 95% CI: 0.79 to 0.85), random forests (sensitivity: 0.72, 95% CI: 0.26 to 0.95, speci