Background: Despite the introduction of Early Warning Scores (EWSs), clinical deterioration (CD) remains an actual problem on the general ward. A next step to counter CD would be to intensify measurement from intermittent 8 hours to continuous measurements. This leads to big data sets of patient monitoring data with great potential. Use of advanced predictive analytics such as Artificial Intelligence (AI), could curb this data into useful alarms. First attempts seem promising, but the evidence remains inconclusive. To create overview of this new research field and to set future perspectives we conducted a scoping review.The primary objective of the scoping review is to identify and specify the state of art models for predicting CD on the general ward. Secondary objective was to identify facilitators, barriers and effects of implementation and use of these models.
Methods: The methodology of Arksey & O’Malley, enhanced Levac et al. was used to perform a scoping review. Data was systematically extracted, charted and summarized. Remarks on implementation and use were identified and categorized as Facilitators, Barriers and Effects using adapted frameworks.
Results: We included 26 studies, divided in 3 groups: using AI, not using AI and sepsis-specific. Only 8 of 26 models were technically implemented of which 2 were clinically used. Input variables differed widely. Output of these models varied from real-time up to a 24 hour prediction for AI and non-AI models and the sepsis models were limited to real-time feedback. Overall AUC’s ranged between .65 and .93 with PPV and NPV ranges of .223-.745 and .930-.999. In sepsis models sensitivity between 51-99% were found compared to 37.3-84.0% for AI and 7.2-52.5% for non-AI, respectively.For the implementation and use of these models 52 facilitators and 42 barriers were identified. Barriers had the upper hand in the technical domain, facilitators dominated the organization domain. Implementation and use of these models lead to 73 reported effects of which 82% were positive.
Conclusions: Advanced predictive analytics shows great potential to counter clinical deterioration on the general ward. Technically the models are feasible to implement, the adaptation of the clinical workflow needs more attention. AUROC-specification of current models are superior to current EWSs in detecting deterioration. Clinical evidence is sporadic and scattered. Consensus on the use of data variables and prediction specifications is yet to be decided upon. Finally and most importantly, uniformity among models is needed to not only compare models, but also to share valuable models and data.