Abstract
The application of machine learning models to big data has become ubiquitous, however their successful translation into clinical practice is currently mostly limited to the field of imaging. Despite much interest and promise, there are many complex and interrelated barriers that exist in clinical settings, which must be addressed systematically in advance of wide-spread adoption of these technologies. There is limited evidence of comprehensive efforts to consider not only their raw performance metrics, but also their effective deployment, particularly in terms of the ways in which they are perceived, used and accepted by clinicians.
The critical care outreach team at St Vincent’s Public Hospital want to automatically prioritise their workload by predicting in-patient deterioration risk, presented as a watch-list application. This work proposes that the proactive management of in-patients at risk of serious deterioration provides a comprehensive case-study in which to understand clinician readiness to adopt deep-learning technology due to the significant known limitations of existing manual processes.
Herein is described the development of a proof of concept application uses as its input the subset of real-time clinical data available in the EMR. This data set has the noteworthy challenge of not including any electronically recorded vital signs data. Despite this, the system meets or exceeds similar benchmark models for predicting in-patient death and unplanned ICU admission, using a recurrent neural network architecture, extended with a novel data-augmentation strategy.
This augmentation method has been re-implemented in the public MIMIC-III data set to confirm its generalisability. The method is notable for its applicability to discrete time-series data. Furthermore, it is rooted in knowledge of how data entry is performed within the clinical record and is therefore not restricted in applicability to a single clinical domain, instead having the potential for wide-ranging impact.
The system was presented to likely end-users to understand their readiness to adopt it into their workflow, using the Technology Adoption Model. In addition to confirming feasibility of predicting risk from this limited data set, this study investigates clinician readiness to adopt artificial intelligence in the critical care setting. This is done with a two-pronged strategy, addressing technical and clinically-focused research questions in parallel.