Anne van Dalen

228 I Chapter 8 more AI applications and studies have been initiated. 19,50,51 The Guangdong Second Provincial General Hospital, for example, plans to incorporate AI image recognition into their infection control system to provide real-time monitoring and an aid for minimizing the risk nosocomial COVID-19 infection. The observing system aims to enhance the sensitivity and accuracy of instant detection in negative pressure isolation wards, which offers creative assistance to combat the COVID-19 outbreak. 50 This application may also be used in the OR to minimize the risk of surgical infection. Indeed, AI in health care has presented some promising and impressive results and is a fertile area of research, as Challen et al. 52 concluded in their review. However, as this study shows the multilingual character of AI in surgery, AI is a complicated and comprehensive field of study. The rapid pace of change, diversity of different techniques, and multiplicity of tuning parameters make it difficult to get a clear picture of how accurate these systems might be in clinical practice or how reproducible they are in different environments. 52 A realistic perspective is needed, balancing the potential for improvement against the risk of negative outcomes. As Yu et al. 8 also concluded, we need to acknowledge the brittleness of these systems, the importance of defining the correct frameworks for their application, and ensure rigorous quality control, including human supervision, to unwanted outcomes. Rigorous prospective trials in a diverse patient population and clinical review of atypical feature statistics are needed, to safeguard the value and coherency of the collected data. 8,52 It is therefore wise to attract knowledge coming from ML experts, ethicists, and lawyers, next to healthcare professionals, to decide on proper fit of use case and safety of AI systems. This study has some limitations to take into account. First, as this is a review study, unpublished data and gray literature, such as technical reports, are not included, which may have strengthened the results. Moreover, the results may have been influenced by a publication bias, especially, because—as this is study shows—AI assistance in the OR is still in its infancy. Park et al. 53 acknowledged the problem of irregular research designs in medical AI studies. This is also displayed by the significant variability in the way results are reported, making it very difficult to combine and compare data across studies. This results in the realization that before any AI tool can be used in clinical practice,