Participants

One hundred and twenty students enrolled in the final two years of a medical degree at Australian universities participated in the study. Australian medical education uses an integrative approach where students learn patient and clinical content throughout their degree. By the final two years of their education, participants would have typically received training in rational and safe prescribing. They also complete the National Prescribing Curriculum, a series of online modules based on the prescribing principles outlined in the World Health Organisation’s Guide to Good Prescribing [23]. Upon completion of these final two years, graduates would begin practice as junior medical officers.

Participants responded to advertisements emailed by medical schools or posted on social media via medical students’ societies. Ethical approval was granted by the ethics committees of Macquarie University and the University of New South Wales. Participants were offered two movie vouchers and a certificate for their participation.

Experiment design

The study had two within-subject factors: quality of CDS (correct, incorrect and no CDS) and task complexity (low, low with interruption and high) providing nine conditions (Fig. 1). Each participant received all nine conditions, completing one scenario in each condition. The experimental control were scenarios presented to participants with no CDS.

The allocation of the nine prescribing scenarios to the nine experimental conditions, the order of presentation, and whether participants received control scenarios first or last were randomized. The position of prescribing and false positive errors in the list of medicines to be prescribed was randomised, allocated at the time of scenario design. The position of alerts was varied depending on the CDS condition that was randomly allocated to the scenario for each participant at the time of enrolment.

Experimental task

Figure 2 provides an example of the participants’ task in this experiment. Participants were presented with nine prescribing scenarios for which they were asked to prescribe medicines using an e-prescribing system. Each scenario presented a brief patient history together with a list of medications to prescribe.

The prescribing scenarios were developed with advice from an expert panel, including four hospital doctors, a medical pharmacology registrar and two pharmacists (including MZR). They were independently reviewed by a consultant physician specialising in pharmacology (RD), to ensure clinical relevance. The scenarios presented hypothetical patient scenarios and involved prescribing tasks that were typical of those undertaken by junior medical officers, based on observations of e-prescribing in a medical ward of a major teaching hospital. A common task performed by junior medical officers is the prescribing of medications using an e-prescribing system upon admission of a patient to hospital, including medicines taken prior to, and those initiated on admission.

Each scenario included one genuine prescribing error, where one of the medicines was clinically contraindicated in that scenario (Additional file 1: Appendix A). These were designed to be unambiguously errors and of sufficient severity in the risk posed to the patient that the medicine should be avoided under all circumstances. To ensure this, the severity of the errors included in the scenarios were independently assessed by a clinical pharmacist (LGP). The error in one scenario was assessed as potentially lethal, five were serious, and three were significant [24]. All other medicines listed in scenarios where carefully chosen so as to be unambiguously free from error.

Scenario complexity was manipulated by varying the amount of information contained in the prescribing scenarios [25]. The nine scenarios were divided into six low-complexity scenarios (each containing six information elements) and three high complexity scenarios (each containing seventeen elements). An information element was classified as either a condition, symptom, test result, prior treatment, allergy, observation, or requested prescription. Each element could potentially interact with other elements in a way that could result in a prescribing error, for example, drug-drug interactions and conditions which may contraindicate the use of a particular medicine. The more elements, the more potential interactions the participant needs to assess. Low-complexity scenarios contained a list of three requested medicines, while high-complexity ones contained eight. The number of elements in each scenario was coded by DL and reviewed by MZR; disagreements were resolved by consensus.

In interruption conditions, participants were interrupted, once per scenario, whilst viewing drug information and presented with a task requiring a response before they could continue. The task required them to seek out and retain in memory three information elements to calculate a dose (Additional file 1: Appendix B).

E-prescribing system

A simulated e-prescribing system was developed which allowed for the manipulation of the triggering and content of CDS alerts. This web-based system was presented to participants as being in development. A medication administration record was not implemented, nor were participants required to specify times of administration.

CDS was provided in the form of alerts (Fig. 3) which were triggered once a prescription was entered. The alert provided a generic warning about the nature of the error, followed by specific details.

Participants could resolve the alert by choosing either to remove (i.e., not prescribe) the medicine or to override the alert with a reason and prescribe that medicine anyway. The alert also provided direct access to drug information for the relevant medicine from the Australian Medicines Handbook [26]. The Australian Medicines Handbook references the Australian formulary and is a gold standard medicines reference. It is evidence-based, reflects Australian best practice and is widely utilised in Australian clinical practice [27]. This reference was also readily accessible from the medication chart and in prescription order entry screens and could be used to identify prescribing errors and verify the information provided by CDS alerts.

Procedure

After having given informed consent, participants completed a pre-experiment questionnaire and watched a brief instructional video on how to use the e-prescribing system. The video included a demonstration of the correct functioning of CDS alerts and how to view drug information.

Participants were instructed as follows: (1) Approach tasks as if they were treating a real patient, exercising all due care; (2) Should they detect any prescribing errors, these should be addressed by not prescribing that medicine; (3) If the error involved an adverse drug interaction between two medicines, only one should be omitted; (4) If there was a discrepancy between CDS and the drug information they should rely on the drug information.

The task was presented as an evaluation of an e-prescribing system under development and participants were told that “Initial testing has shown that alerts are highly accurate, but occasionally have been incorrect. Therefore, you should always double check with the inbuilt drug information reference.” No information was provided on what types of errors the system would check and alert. Once all scenarios were completed, participants completed a post-experiment questionnaire and were then debriefed.

Outcome measures

Prescribing errors were classified according to the definitions of prescribing error categories provided by Westbrook et al. [28]. The potential severity of prescribing errors was assessed by a clinical pharmacist (LGP) using the severity error classification scheme described in Dornan et al. [24].

Statistical analyses

The presence of AB was tested using McNemar’s test [29] comparing errors between scenarios with incorrect CDS and scenarios with no CDS (control). It was estimated that 120 participants would be required to detect a 25% or greater difference (two-tailed) in errors between the control and incorrect CDS scenarios with 80% power and p < 0.05 [30]. With five hypotheses tested, a Bonferroni correction was applied to control for the increased risk of making a Type I error when testing multiple hypotheses [31]. With the desired alpha of 0.05, the corrected alpha against which all significance probabilities were evaluated became 0.01. Significance probabilities are only reported for comparisons between individual conditions, but not for aggregate figures by quality of CDS, which include multiple observations from each participant. Scenarios in which participants did not experience an interruption were excluded from the interruption analysis (n = 36 with no CDS, n = 22 with correct CDS, and n = 4 with incorrect CDS).

Correct CDS decreased prescribing errors

However, correct CDS did not significantly alter commission errors (Table 2). This was the case for low complexity (p = 1.0, n = 120), low + interrupt (p = .219, n = 65) and high complexity scenarios (p = .678, n = 120). Participants also made omission errors by overriding correct CDS alerts in 8.3% of scenarios and commission errors by not prescribing the safe, comparator, medicines in 5.3% of scenarios.

Incorrect CDS increased prescribing errors

Participants missed 28.7% more genuine prescribing errors (omission errors) when assisted by incorrect CDS compared to no CDS (Table 2). These differences were statistically significant across all levels of complexity, with 33.3% more errors in low complexity (p < .0001, n = 120), 24.5% more errors in low + interrupt (p = .009, n = 82) and 26.7% more errors in high complexity scenarios (p < .0001, n = 120).

Overall participants made 56.9% more commission errors (did not prescribe safe medicines) when they received false positive alerts from incorrect CDS compared to when they received no CDS (Table 2). These differences were statistically significant across all levels of complexity, with participants in scenarios receiving false positive alerts making 65.8% more errors in low complexity (p < .0001, n = 120), 53.5% more errors in low + interrupt (p < .0001, n = 82) and 51.7% more errors in high complexity scenarios (p < .0001, n = 120).

Interruptions to prescribing and scenario complexity did not impact automation bias

Interruptions did not affect omission or commission errors, or errors in the control scenarios. In interrupted scenarios with incorrect CDS there were 0.1% more omission errors (p = 1.0 n = 116) and 9.7% fewer commission errors (p = .08, n = 116). In interrupted control scenarios there were 8.9% more omission errors (p = .2, n = 84) and 2.6% more commission errors (p = .22, n = 84). All of these were non-significant.

Scenario complexity did not affect omission or commission errors, or errors in the control scenarios. In high complexity scenarios with incorrect CDS there were 4.2% fewer omission errors (p = .46, n = 120) and 5.0% fewer commission errors (p = .35, n = 120). In high complexity control scenarios there were 2.5% more omission errors (p = .75, n = 120) and 9.2% more commission errors (p = .007, n = 120). The only significant difference was between low and high complexity control scenarios.

More omission than commission errors

Overall participants made 13.5% more omission than commission errors when provided with incorrect CDS, however, this was only significant in the low complexity + interrupt condition, all others were non-significant. There were 10% more omission errors in low complexity (p = .065, n = 120), 19.8% more in low + interrupt (p = .001, n = 116) and 10.8% more in high complexity scenarios (p = .079, n = 120).

Main findings

This is the first study to find evidence of automation bias in the presence of e-prescribing CDS alerts. We found that when CDS was correct it reduced overall prescribing errors by 58.8%. This is consistent with prior literature showing that e-prescribing CDS can reduce prescribing errors [3–5]. However, when CDS was incorrect it increased errors by 86.6%. This increase was due to AB, that is, the ability of incorrect CDS to adversely influence participant prescribing decisions.

We found evidence of participants making omission errors, by failing to detect 28.7% more prescribing errors when CDS failed to provide alerts, compared to a control condition with no CDS. This finding was significant across all levels of task complexity and is potentially serious as the missed prescribing errors were classified as being of significant to potentially lethal severity, with most classified as serious severity.

Likewise, participants made commission errors, acting on clinically incorrect, false positive alerts, by not prescribing 56.9% more necessary medicines compared to the control condition. This was significant across all levels of task complexity.

These findings are consistent with and add to the research on automation bias in healthcare. Finding evidence of omission errors in the computer-aided detection of cancers in screening mammography [15, 16] and commission errors in the computerized interpretation of EKGs, [17] answering clinical questions assisted by CDS, [18] and deciding what to prescribe for clinical scenarios [19].

Interestingly, while participants were found to over-rely on automation, there was evidence of disagreement with the CDS provided to them. Participants’ overrode correct alerts and in doing so made prescribing errors which CDS was warning them to avoid. They also did not prescribe medicines which did not contain errors and for which there were no alerts. Reasons provided for overriding correct CDS alerts commonly referred to the condition for which the medicine was intended to treat (e.g. “VTE risk and pain management”, “vomiting”) or indicated that the medicine was regularly taken by the patient (e.g. “patient usual dose”). Participants commonly cited the lack of a true contraindication as the reason for overriding incorrect CDS alerts with many referring to the drug information. For example, “There is not any interaction listed on the drug information”. However, regular patient medicines and the condition treated were also mentioned as reasons for overriding incorrect CDS alerts. This suggests that not only did participants have trouble determining when CDS was wrong, but some also had trouble recognizing when it was right and that the alerts, or lack thereof, were beneficial and should be heeded.

Implications

When clinical decision support is right, it can reduce prescribing errors by providing an important opportunity to detect and recover from prescribing errors. However, the finding of automation bias suggests that this additional layer of defence weakens or, at worst, becomes a replacement for the clinician’s own efforts in error detection with error detection delegated to CDS, without adequate oversight.

An intuitive solution to the problem of AB is to produce CDS systems that are less prone to error. While this may reduce the overall error rate, highly accurate automation is known to increase the rate of AB [25]. In other words, when automation does fail, the clinician will be even less able to detect it.

A key problem is that users seem to have difficulty in determining when CDS should and should not be relied on. Indeed, human factors research reports an inverse relationship between measures of verification, such as viewing drug references, and AB commission errors [10, 11]. So far, interventions to counter AB have had little success [37–39]. These include a number specifically targeted at verification, such as exposure to automation failures; [10] training about AB; and providing prompts to verify [40]. Compounding this problem further are findings of a looking-but-not-seeing effect or inattentional blindness where participants have made AB errors despite accessing sufficient information to assess that automation was incorrect [12, 13].

Verification, the means by which a user can determine whether the CDS they receive is correct, is key to the mitigation of AB. However, the lack of successful interventions indicates that more research is needed on how to best assist users with this crucial task.

This study has established that there is a risk of automation bias in electronic prescribing with senior medical students, who will soon be entering clinical practice as junior doctors. In doing this, we have also demonstrated a methodology for detecting AB in e-prescribing. The true rates and effects of AB in working clinical settings will require further studies and indeed is likely to vary by clinicians’ experience and familiarity with medications, clinical setting, patient complexity, and the particular decision support system used. All this is future work. Likewise, the lack of an effect of task complexity, even in control conditions, was surprising and something future studies will need to address. This might be achieved by varying clinician experience with prescribing and e-prescribing systems. Complexity could also incorporate familiarity with medicines, varying between simple, commonly-used to complex, rarely-prescribed regimes.

Clinicians need to be mindful that CDS can and does fail [6]. Ideally, clinicians should make every effort to detect prescribing errors, allowing CDS to function as an independent check for errors rather than relying on it as a replacement of their own error detection efforts.

Limitations

Several limitations arise from the design of this study. While participants were instructed to approach the task as if they were treating a real patient, exercising all due care, the prescribing task was simulated, and prescribing errors were without consequence.

Also as an experiment, we cannot make any inferences about the true effect size or rate of AB in clinical settings as this will vary with, the user, the tasks being performed and the accuracy of the decision support provided. Likewise, the nature and incidence of the provided opportunities for prescribing errors may not be representative of those encountered in clinical practice.

The lack of a difference in prescribing errors between the low and high complexity control scenarios limited our ability to assess the impact of task complexity on AB.

Finally, the use of medical students with little experience in both prescribing medicines and using e-prescribing systems provides an indication of how CDS will impact new clinicians entering practice but limits generalisability for experienced prescribers or clinicians with e-prescribing experience.