Are AI-based surveillance systems for healthcare-associated infections ready for clinical practice? A systematic review and meta-analysis.
Journal:
Artificial intelligence in medicine
PMID:
40286586
Abstract
Healthcare-associated infections (HAIs) are a global public health concern, imposing significant clinical and financial burdens. Despite advancements, surveillance methods remain largely manual and resource-intensive, often leading to underreporting. In this context, automation, particularly through Artificial Intelligence (AI), shows promise in optimizing clinical workflows. However, adoption challenges persist. This study aims to evaluate the current performance and impact of AI in HAI surveillance, considering technical, clinical, and implementation aspects. We conducted a systematic review of Scopus and Embase databases following PRISMA guidelines. AI-based models' performances, accuracy, AUC, sensitivity, and specificity, were pooled using a random-effect model, stratifying by detected HAI type. Our study protocol was registered in PROSPERO (CRD42024524497). Of 2834 identified citations, 249 studies were reviewed. The performances of AI models were generally high but with significant heterogeneity between HAI types. Overall pooled sensitivity, specificity, AUC, and accuracy were respectively 0.835, 0.899, 0.864, and 0.880. About 35.7 % of studies compared AI system performance with alternative automated or standard-of-care surveillance methods, with most achieving better or comparable results to clinical scores or manual surveillance. <7.6 % explicitly measured AI impact in terms of improved patient outcomes, workload reduction, and cost savings, with the majority finding benefits. Only 30 studies deployed the model in a user-friendly tool, and 9 tested it in real clinical practice. In this systematic review, AI shows promising performance in HAI surveillance, although its routine application in clinical practice remains uncommon. Despite over a decade, retrieved studies offer scant evidence on reducing burden, costs, and resource use. This prevents their potential superiority over traditional or simpler automated surveillance systems from being fully evaluated. Further research is necessary to assess impact, enhance interpretability, and ensure reproducibility.