Automation bias and prescribing decision support – rates, mediators and mitigators

Purpose: Computerised clinical decision support systems (CDSS) are implemented within healthcare settings as a method to improve clinical decision quality, safety and effectiveness, and ultimately patient outcomes. Though CDSSs tend to improve practitioner performance and clinical outcomes, relatively little is known about specific impact of inaccurate CDSS output on clinicians. Although there is high heterogeneity between CDSS types and studies, reviews of the ability of CDSS to prevent medication errors through incorrect decisions have generally been consistently positive, working by improving clinical judgement and decision making. However, it is known that the occasional incorrect advice given may tempt users to reverse a correct decision, and thus introduce new errors. These systematic errors can stem from Automation Bias (AB), an effect which has had little investigation within the healthcare field, where users have a tendency to use automated advice heuristically. Research is required to assess the rate of AB, identify factors and situations involved in overreliance and propose says to mitigate risk and refine the appropriate usage of CDSS; this can provide information to promote awareness of the effect, and ensure the maximisation of the impact of benefits gained from the implementation of CDSS. Background: A broader literature review was carried out coupled with a systematic review of studies investigating the impact of automated decision support on user decisions over various clinical and non-clinical domains. This aimed to identify gaps in the literature and build an evidence-based model of reliance on Decision Support Systems (DSS), particularly a bias towards over-using automation. The literature review and systematic review revealed a number of postulates - that CDSS are socio-technical systems, and that factors involved in CDSS misuse can vary from overarching social or cultural factors, individual cognitive variables to more specific technology design issues. However, the systematic review revealed there is a paucity of deliberate empirical evidence for this effect. The reviews identified the variables involved in automation bias to develop a conceptual model of overreliance, the initial development of an ontology for AB, and ultimately inform an empirical study to investigate persuasive potential factors involved: task difficulty, time pressure, CDSS trust, decision confidence, CDSS experience and clinical experience. The domain of primary care prescribing was chosen within which to carry out an empirical study, due to the evidence supporting CDSS usefulness in prescribing, and the high rate of prescribing error. Empirical Study Methodology: Twenty simulated prescribing scenarios with associated correct and incorrect answers were developed and validated by prescribing experts. An online Clinical Decision Support Simulator was used to display scenarios to users. NHS General Practitioners (GPs) were contacted via emails through associates of the Centre for Health Informatics, and through a healthcare mailing list company. Twenty-six GPs participated in the empirical study. The study was designed so each participant viewed and gave prescriptions for 20 prescribing scenarios, 10 coded as “hard” and 10 coded as “medium” prescribing scenarios (N = 520 prescribing cases were answered overall). Scenarios were accompanied by correct advice 70% of the time, and incorrect advice 30% of the time (in equal proportions in either task difficulty condition). Both the order of scenario presentation and the correct/incorrect nature of advice were randomised to prevent order effects. The planned time pressure condition was dropped due to low response rate. Results: To compare with previous literature which took overall decisions into account, taking individual cases into account (N=520), the pre advice accuracy rate of the clinicians was 50.4%, which improved to 58.3% post advice. The CDSS improved the decision accuracy in 13.1% of prescribing cases. The rate of AB, as measured by decision switches from correct pre advice, to incorrect post advice was 5.2% of all cases at a CDSS accuracy rate of 70% - leading to a net improvement of 8%. However, the above by-case type of analysis may not enable generalisation of results (but illustrates rates in this specific situation); individual participant differences must be taken into account. By participant (N = 26) when advice was correct, decisions were more likely to be switched to a correct prescription, when advice was incorrect decisions were more likely to be switched to an incorrect prescription. There was a significant correlation between decision switching and AB error. By participant, more immediate factors such as trust in the specific CDSS, decision confidence, and task difficulty influenced rate of decision switching. Lower clinical experience was associated with more decision switching (but not higher AB rate). The rate of AB was somewhat problematic to analyse due to low number of instances – the effect could potentially have been greater. The between subjects effect of time pressure could not be investigated due to low response rate. Age, DSS experience and trust in CDSS generally were not significantly associated with decision switching. Conclusion: There is a gap in the current literature investigating inappropriate CDSS use, but the general literature supports an interactive multi-factorial aetiology for automation misuse. Automation bias is a consistent effect with various potential direct and indirect causal factors. It may be mitigated by altering advice characteristics to aid clinicians’ awareness of advice correctness and support their own informed judgement – this needs further empirical investigation. Users’ own clinical judgement must always be maintained, and systems should not be followed unquestioningly.