Artificial Intelligence-based tools to control healthcare associated infections: where do we stand

Abstract

Abstract Background Control of Healthcare associated infections (HAI) is a key public health concern in Europe. Current HAI surveillance systems are based on manual medical records review, vulnerable to misclassification & expensive. Artificial intelligence (AI), as a digital tool, offers great potential to HAI control. Still, scant evidence is available on both its practice and impact. Methods As part of a broader multidisciplinary project, we conducted a systematic review to retrieve, pool and critically apprize all the available evidence on practice, performance and impact of AI-based HAI control programmes. We followed PRISMA and searched the Medline and Embase databases for relevant studies. Included studies were stratified by HAI type and outcomes of interest, including all possible performance measures, clinical, organizational and economic outcomes. Results We screened 2873 records, resulting in 27 papers included in the review. Studies were carried out in 9 countries, the majority in the US (56%), 18.5% in EU countries, 25.9% published in 2018. Two thirds of studies focused on selected types of infections. Study designs were very diverse and performance observed for HAI detection were very heterogeneous, precluding pooled calculation of summary diagnostic accuracy estimates in most instances, but generally higher than non AI-based models. Overall performance measures of AI algorithms were: sensitivity range 19%-92%, accuracy 70.2%-96.1%. Conclusions Use of AI for HAI surveillance of HAI has increased reliability compared to traditional surveillance or to automated surveillance models. With ongoing improvements in information technology, implementation of AI models will improve the quality and capacity of surveillance will support hospital HAI surveillance. Main messages AI offer great potential to healthcare associated infections control. Preliminary evidence show AI-based models have perform better than manual or automated models for HAIs detection.