Skip to main content
. 2021 Mar 19;4:54. doi: 10.1038/s41746-021-00423-6

Table 3.

Evaluation of the potential of artificial intelligence to improve patient safety in the eight harm domains.

Patient safety domain Likelihood of impact
Healthcare-associated infections AI may have a moderate impact on the reduction of HAIs given that current evidence-based practices are already effective when applied well.
Adverse drug events AI can play a major role in ADE prevention. As more patients at risk of ADEs are accurately identified before a medication is administered or prescribed, a greater proportion of these events will become preventable. However, a key challenge lies in the lack of integrated high-quality datasets in which ADEs have been accurately captured. A variety of automated approaches have been effective at identifying patients likely to have experienced an ADE, but typically clinician adjudication is still required. ML could also help identify patients who may benefit from additional testing for specific single nucleotide polymorphisms to guide optimal drug therapy. These methods may also help to identify signals from the remainder of the genome beyond single nucleotide polymorphisms which may have prognostic impact.
Venous thromboembolism We believe that AI will have a moderate effect on the reduction of VTE, as current evidence-based preventive strategies are already effective. AI solutions could provide further insights by identifying patients who could benefit from diagnostic testing for inherited thrombotic disorders to inform management of their condition.
Surgical complications AI can be expected to have a moderate impact on the prediction and prevention of surgical complications both in the operating room and during recovery. Most complications felt to be preventable today are related to delayed diagnoses or intervention, technical issues, and infections. Given the overlap with other harm domains, focusing on advances in these other areas will likely also improve surgical safety.
Pressure ulcers Pressure ulcers represent an attractive target with moderate to high potential for AI to prevent harm. Novel data sources such as motion and fluid sensors are now available, and large numbers of traditional clinical variables can be combined with the sensing data to predict who is at risk to guide evidence-based prevention.
Falls AI is anticipated to have a moderate impact on fall prevention, given that this area has already received substantial attention and current risk mitigation strategies are effective. As with pressure ulcers, clinical data combined with sensing data can be used to predict when falls may occur, and which ones are likely to be associated with the most harm.
Decompensation Leveraging novel data sources and AI has high potential to improve the prediction of decompensation to guide preventive strategies as well as early intervention to mitigate the impacts including premature death, given that current approaches are not effective. Given the serious nature of these events, preventing decompensation is a particularly attractive target. ML can deeply analyze data, beyond the standard values of heart rate or heart rate variability and will be critical to improving detection of decompensation and subsequent intervention.
Diagnostic errors Diagnostic error is the most complex of the eight harm domains with vast opportunities for improvement using novel data sources and AI. ML could help to reduce the frequency of diagnostic errors by leveraging pattern recognition, bias minimization, and infinite capacity, areas where diagnosticians often falter. Although this area has garnered a lot of attention, many outstanding challenges remain, and current solutions only address a small fraction of what is possible. Most crucial to constructing valuable ML algorithms that help to reduce diagnostic error is the availability of large databases that accurately report errors.

ADE adverse drug event, AI artificial intelligence, HAI healthcare-associated infection, ML machine learning, VTE venous thromboembolism.