Automation bias in electronic prescribing: the effects of over-reliance on clinical decision support in relation to errors, cognitive load and verification

Prescribing errors are a leading preventable cause of patient harm. Clinical decision support (CDS) can improve safety by alerting clinicians to potential errors as they enter orders into e-prescribing systems. However, this can introduce the risk of automation bias; clinicians may over-rely on CDS, thereby reducing vigilance in information seeking and processing. Problematically, CDS may not detect all significant errors or may generate alerts which are not clinically significant. Omission errors occur when clinicians fail to detect prescribing errors because they were not alerted, and commission errors occur where incorrect advice is wrongly acted upon. To date, there has been little research on automation bias in healthcare, where tasks, decision support and task complexity are likely to differ from those utilised in existing research which comes mostly from the heavily automated domains of aviation and process control. This thesis examines the risk of automation bias in e-prescribing that is assisted by CDS and whether this risk is mediated by task complexity. It also examines the relationship between automation bias errors, cognitive load, and verification of CDS. One hundred and twenty students in the final two years of a medical degree prescribed medicines for nine clinical scenarios using a simulated e-prescribing system in a randomised controlled experiment. The quality of CDS (correct, incorrect and no CDS) and task complexity (low, low with interruption and high) were varied within-subjects. Omission errors (failure to detect prescribing errors), commission errors (acceptance of false positive alerts), cognitive load, and verification of CDS (access of drug references) were measured. Errors. Compared to no CDS, incorrect CDS significantly increased omission errors by 33.3% (p < .0001), 24.5% (p = .009), and 26.7% (p < .0001) and commission errors by 65.8% (p < .0001), 53.5% (p < .0001), and 51.7% (p < .0001), for low-, low- with interruption and high-complexity scenarios, respectively. Task complexity and interruptions did not affect errors. Cognitive Load. The use of CDS reduced cognitive load in high complexity conditions compared to no CDS, F(2,117)=4.72,p=.015. Omission errors were associated with significantly lower cognitive load with incorrect and no CDS, F(1,636.49)=3.79,p=.023. Verification. Lower view times (as a percentage of task time) increased omission errors, F(3, 361.914)=4.498, p=.035, and commission errors, F(1, 346.223)=2.712, p=.045. View times were lower in CDS-assisted compared to unassisted conditions, F(2, 335.743)=10.443, p<.001. This thesis contributes the first evidence of automation bias in e-prescribing, a common clinical decision-making task aided by a frequently encountered form of CDS. It also contributes the first evidence of the relationship between automation bias and reduced allocation of cognitive resources. Participants made omission errors by failing to detect prescribing errors not alerted by CDS and made commission errors by accepting incorrect false-positive alerts. The presence of CDS reduced cognitive load and verification, and increased errors when CDS was incorrect. These effects were exacerbated under conditions of high task complexity, suggesting high complexity may be a risk factor. Curiously, however, task complexity had no effect on errors. Participants who made automation bias errors allocated fewer cognitive resources and verified less than those who avoided errors. These findings support the cognitive miser hypothesis of automation bias that CDS alerts were used as a heuristic or mental shortcut for detecting and avoiding prescribing errors. It is highly likely that when clinicians suffer an automation bias, they reduce both verification behaviours and the cognitive resources allocated to processing information. This, in turn, compromises their ability to detect problems, which could potentially result in patient harm. The challenge is to foster appropriate reliance on CDS, which improves efficiency and reduces errors when correct but can lead to automation bias errors when incorrect. Verification of CDS provides a key means to discriminate correct from incorrect CDS that could prevent automation bias errors. More research will be needed on how to best assist clinicians with this crucial task whilst simultaneously leveraging the enhanced efficiency and safety offered by correct CDS. Clinicians should be mindful of the limitations of CDS and the possibility that it can fail. They should be ever-vigilant and ready to verify whenever unfamiliarity or uncertainty is present, or a risk of patient harm is suspected.