Stroke physicians’ and staff perspectives on machine learning to optimise thrombolysis decision making in stroke: a qualitative study

Abstract

BACKGROUNDSAMueL-2 (Stroke Audit for Machine Learning Project) working with the Sentinel Stroke National Audit Programme (SSNAP) developed clinical pathway and machine learning computer models to investigate variation in thrombolysis use. We investigated how this modelling could be designed and adapted to inform clinical practice and support optimal implementation of thrombolysis by exploring the perspectives of physicians and other staff whose work relates to acute stroke care. RESEARCH QUESTIONWhat should a machine-learning model based on SSNAP data look like, do, and deliver if it is to optimise improvement, and reduce unwarranted variation, in thrombolysis?OBJECTIVES1. To generate empirically and theoretically informed knowledge about how thrombolysis is currently delivered, centred on physicians’ views, understandings, and practices.2. To learn more about how stroke physicians’ and staff think and feel about or use SSNAP, and about the use of machine learning in improving clinical practice.DESIGN AND METHODS We used focussed observations, semi-structured interviews and documentary analysis, to examine perceptions of thrombolysis, SSNAP, and machine learning. The Non-adoption, Abandonment, and Challenges to the Scale-Up, Spread, and Sustainability of Health and Care Technologies (NASSS) framework was used as a sensitising device to help us understand socio-technical factors likely to affect adoption and scale-up of SAMueL-2 technology.SETTING AND PARTICIPANTSHospitals were purposefully selected due to low rates of thrombolysis, differing stroke pathways, and for geographical variation. We conducted 184 hours of focussed observation in three NHS Trusts in England; comprising day/evening/night and weekend shifts and relevant meetings. We also observed online meetings of Integrated Stroke Delivery Networks (ISDNs) and other organisations with strategic overview of stroke services. 20 participants from the three observation sites and five key informants from other sites took part in semi-structured interviews.RESULTS We present findings in relation to six NASSS domains: the condition, the technology, the value proposition, the intended adopters, the healthcare organisation, and the wider system. Our data showed participants were hopeful the SAMueL-2 technology could address variance in thrombolysis practice. It was seen as particularly suitable for junior clinicians, non-stroke specialists and at district general hospitals and offered value for training, reviewing clinical cases, and quality improvement.LIMITATIONS Our exploratory findings are not transferrable to all staff involved in acute stroke care/administration. CONCLUSIONS We identified three key learning points. First, given reservations expressed about SSNAP data, it is important to reassure intended adopters about the integrity of modelling based on this data. Second, evidence indicated ED physicians may have less confidence in the evidence base for thrombolysis. More work needs to be done with the ED physician community to build trust in the SAMuel-2 technology: recruiting ED physicians as brokers/clinical champions may address this. Third, perceived lack of funding and stroke workforce shortages may impede quality improvement and adoption of new technologies such as SAMueL-2. These concerns must be addressed to ensure sustained use and adoption. The next phase of the research will focus on the seventh NASSS domain relating to embedding and adaptation of the technology over time.