A recent Institute of Medicine report has brought the problem of medical error under intense scrutiny[1]. While the use of computerized prescription software has been shown to substantially reduce the incidence of medication-related error [2, 3], few solutions have demonstrated a similar impact on diagnostic error. Diagnostic errors impose a significant burden on modern healthcare: they account for a large proportion of medical adverse events in general [4–6], and form the second leading cause for malpractice suits against hospitals [7]. In particular, diagnostic errors of omission (DEO) during acute medical assessment, resulting from cognitive biases such as 'premature closure' and 'confirmation bias', lead to incomplete diagnostic workup and 'missed diagnoses' [8]. This is especially relevant in settings such as family practice [9], as well as hospital areas such as the emergency room and critical care [11, 12], 20% of patients discharged from emergency rooms raised concerns in a recent survey that their clinical assessment had been complicated by diagnostic error [12]. The use of clinical decision-support systems (DSS) has been one of many strategies proposed for the reduction of diagnostic errors in practice [13]. Consequently, a number of DSS have been developed over the past few years to assist clinicians during the process of medical diagnosis [14–16].

Even though studies of several diagnostic DSS have demonstrated improved physician performance in simulated (and rarely real) patient encounters [17, 18], two specific characteristics may have contributed to their infrequent use in routine practice: intended purpose and design. Many general diagnostic DSS were built as 'expert systems' to solve diagnostic conundrums and provide the correct diagnosis during a 'clinical dead-end' [19]. Since true diagnostic dilemmas are rare in practice [20], and the initiative for DSS use had to originate from the physician, diagnostic advice was not sought routinely, particularly since clinicians prefer to store the patterns needed to solve medical problems in their heads [21]. There is, however, evidence that clinicians frequently underestimate their need for diagnostic assistance, and that the perception of diagnostic difficulty does not correlate with their clinical performance [22]. In addition, due to the demands of the information era [23]. diagnostic errors may not be restricted to cases perceived as being difficult, and might occur even when dealing with common problems in a stressful environment under time pressure [24]. Further, most 'expert systems' utilized a design in which clinical data entry was achieved through a controlled vocabulary specific to each DSS. This process frequently took > 15 minutes, contributing to infrequent use in a busy clinical environment [25]. These 'expert systems' also provided between 20 and 30 diagnostic possibilities [26], with detailed explanations, leading to a lengthy DSS consultation process.

In order to significantly affect the occurrence of diagnostic error, it seems reasonable to conclude that DSS advice must therefore be readily available, and sought, during most clinical encounters, even if the perceived need for diagnostic assistance is minor. Ideally, real-time advice for diagnosis can be actively provided by integrating a diagnostic DSS into an existing electronic medical record (EMR), as has been attempted in the past [27, 28]. However, the limited uptake of EMRs capable of recording sufficient narrative clinical detail currently in clinical practice indicates that a stand-alone system may prove much more practical in the medium term [29]. The key characteristic of a successful system would be the ability to deliver reliable diagnostic reminders rapidly following a brief data entry process in most clinical situations. ISABEL (ISABEL Healthcare, UK) is a novel Web-based pediatric diagnostic reminder system that suggests important diagnoses during clinical assessment [30, 31]. The development of the system and its underlying structure have been described in detail previously [32, 33]. The main hypotheses underlying the development of ISABEL were that the provision of diagnostic reminders generated following a brief data entry session in free text would promote user uptake, and lead to improvement in the quality of diagnostic decision making in acute medical settings. The reminders provided (a set of 10 in the first instance) aimed to remind clinicians of important diagnoses that they might have missed in the workup. Data entry is by means of natural language descriptions of the patient's clinical features, including any combination of symptoms, signs and test results. The system's knowledge base consists of natural language text descriptions of > 5000 diseases, in contrast to most 'expert systems' that use complex disease databases [34–36]. The advantages and trade-offs of these differences in system design have been discussed in detail elsewhere [37]. In summary, although the ability to rapidly enter patient features in natural language to derive a short-list of diagnostic suggestions may allow frequent use by clinicians during most patient encounters, variability resulting from the use of natural language for data entry, and the absence of probability ranking, may compromise the accuracy and usefulness of the diagnostic suggestions.

The overall evaluation of the ISABEL system was planned in systematic fashion in a series of consecutive studies [38].

a) An initial clinical performance evaluation: This would evaluate the feasibility of providing relevant diagnostic suggestions for a range of cases when data is entered in natural language. System accuracy, speed and relevance of suggestions were studied.

b) An assessment of the impact of the system in a quasi-experimental setting: This would examine the effects of diagnostic decision support on subjects using simulated cases.

c) An assessment of the impact of the system in a real life setting: This would examine the effects of diagnostic advice on clinicians in real patients in their natural environment.

In the initial performance evaluation, the ISABEL system formed the unit of intervention, and the quality of its diagnostic suggestions was validated against data drawn from 99 hypothetical cases and 100 real patients. Key findings from cases were entered into the system in free text by one of the developers. The system included the final diagnosis in 95% of the cases [39]. This design was similar to early evaluations of a number of other individual diagnostic DSS [40–42], as well as a large study assessing the performance characteristics of four expert diagnostic systems [43]. Since this step studied ISABEL in isolation, and did not include users uninvolved in the development of the system, it was vital to examine DSS impact on decision making by demonstrating in the subsequent step that the clinician-DSS combination functioned better than either the clinician or the system working in isolation [44, 45]. Evaluation of impact is especially relevant to ISABEL: despite good system performance when tested in isolation, clinicians may not benefit from its advice either due to variability associated with user data entry leading to poor results, or the inability to distinguish between diagnostic suggestions due to the lack of ranking [46]. A previous evaluation of Quick Medical Reference (QMR) assessed a group of clinicians working a set of difficult cases, and suggested that the extent of benefit gained by different users varied with their level of experience [47].

In this study, we aimed to perform an impact evaluation of ISABEL in a quasi-experimental setting in order to quantify the effects of diagnostic advice on the quality of clinical decisions made by various grades of clinicians during acute assessment, using a mix of easy and difficult simulated cases drawn from all pediatric sub-specialties. Study design was based on an earlier evaluation of the impact of ILIAD and QMR on diagnostic reasoning in a simulated environment [48]. Our key outcome measure focused on appropriateness of decisions during diagnostic workup rather than accuracy in identifying the correct diagnosis. The validity of textual case simulations has previously been demonstrated in medical education exercises [49], and during the assessment of mock clinical decision making [50, 51].

The simulated field study involved recording subjects' clinical decisions regarding diagnoses, test-ordering and treatment for a set of simulated cases, both before and immediately after DSS consultation. The impact of diagnostic reminders was determined by measuring changes in the quality of decisions made by subjects. In this study, the quality of ISABEL's diagnostic suggestion list per se was not examined. The study was coordinated at Imperial College School of Medicine, St Mary's Hospital, London, UK between February and August 2002. The study was approved by the Local Research Ethics Committee.

Study procedure

Subjects were allowed to complete their assessments of the simulated cases from any computer connected to the Internet at any time (i.e. they were not supervised). After logging into the trial website, subjects were presented with text from a case simulation. They assessed the case, abstracted the salient clinical features according to their own interpretation of the case, and entered them into the designated search query box in free text. Following this, they keyed in their decisions regarding diagnostic workup, test-ordering and treatment into the designated textboxes. These constituted pre-DSS clinical decisions. See figure 1 for an illustration of this step of the study procedure. On submitting this information, a list of diagnostic suggestions was instantly presented to the subject based on the abstracted clinical features. The subjects could not read the case text again at this stage, preventing them from processing the case a second time, thus avoiding 'second-look' bias. Diagnostic suggestions were different for different users since the search query was unique for each subject, depending on their understanding of the case and how they expressed it in natural language. On the basis of the diagnostic suggestions, subjects could modify their pre-DSS clinical decisions by adding or deleting items: these constituted post-DSS clinical decisions. All clinical decisions, and the time taken to complete each step, were recorded automatically. See figure 2 for an illustration of this step of the study procedure. The text from one case, and the variability associated with its interpretation during the study, is depicted in figure 3.

Each subject was presented with 12 cases such that one of the pair drawn from each sub-specialty was displayed. Cases were presented in random order (in no particular order of sub-specialty). Subjects could terminate their session at any time and return to complete the remainder of cases. If a session was terminated midway through a case, that case was presented again on the subject's return. If the website detected no activity for > 2 hours, the subject was automatically logged off, and the session was continued on their return. All subjects had 3 weeks to complete their assigned 12 cases. Since each case was used more than once, by different subjects, we termed each attempt by a subject at a case as a 'case episode'.

Scoring metrics

We aimed to assess if the provision of key diagnostic reminders would reduce errors of omission in the simulated environment. For the purposes of this study, a subject was defined to have committed a DEO for a case episode if they failed to include all 'clinically important diagnoses' in the diagnostic workup (rather than failing to include the 'correct diagnosis'). A diagnosis was judged 'clinically important' if an expert panel working the case independently decided that the particular diagnosis had to be included in the workup in order to ensure safe and appropriate clinical decision making, i.e. they would significantly affect patient management and/or course, and failure to do so would be construed clinically inadequate. The expert panel comprised two general pediatricians with > 3 years consultant level experience. 'Clinically important' diagnoses suggested by the panel thus included the 'most likely diagnosis/es' and other key diagnoses; they did not constitute a full differential containing all plausible diagnoses. This outcome variable was quantified by a binary measure (for each case episode, a subject either committed a DEO or not). Errors of omission were defined for tests and treatments in similar fashion.

We also sought a more sensitive assessment of changes in the quality of clinical decisions made by subjects in this study. Since an appropriate and validated instrument was essential for this purpose, a measurement study was first undertaken to develop and validate such an instrument. The measurement study, including the development and validation of a diagnostic quality score and a management plan quality score, has previously been reported in detail [52]. The scoring process, tested using a subset of cases worked on by clinicians during this study (190 case episodes), was reliable (intraclass correlation coefficient 0.79) and valid (face, construct and concurrent validity). During the scoring process, the expert panel was provided an aggregate list of decisions drawn from all subjects (pre- and post-DDSS consultation) for each case. They provided measurements of quality for each of the clinical decisions in addition to identifying 'clinically important' decisions for each case. Prior to scoring, one investigator (PR) mapped diagnoses proposed by subjects and the expert panel to the nearest equivalent diagnoses in the ISABEL database. Quality of each diagnostic decision was scored for the degree of plausibility, likelihood in the clinical setting, and its impact on further patient management. These measurements were used to derive scores for each set of subjects' clinical decisions (diagnostic workup, tests and treatments). As per the scoring system, subjects' decision plans were awarded the highest score (score range: 0 to 1) only if they were both comprehensive (contained all important clinical decisions), and focused (contained only important decisions). Scores were calculated for each subject's diagnostic, test-ordering and treatment plans both pre- and post-ISABEL consultation. Figure 4 provides a schematic diagram of the complete scoring procedure.

We have shown in this study that errors of omission occur frequently during diagnostic workup in an experimental setting, including in cases perceived as being common in routine practice. Such errors seem to occur in most subjects, irrespective of their level of experience. We have also demonstrated that it is possible to influence clinicians' diagnostic workup and reduce errors of omission using a stand-alone diagnostic reminder system. Following DSS consultation, the quality of diagnostic, test-ordering and treatment decisions made by various grades of clinicians improved for a range of cases, such that a clinically important alteration in diagnostic decision-making resulted in 12.5% of all consultations (1 in 8 episodes of system use).

In a previous study assessing the impact of ILIAD and QMR, in which only diagnostically challenging cases were used in an experimental setting, Friedman et al showed that the 'correct diagnosis' was prompted by DSS use in approximately 1 in 16 consultations [48]. Although we used a similar experimental design, we used a mix of easy as well as difficult cases to test the hypothesis that incomplete workup was encountered in diagnostic conundrums as well as routine clinical problems. Previous evaluations of expert systems used the presence of the 'correct' diagnosis as the main outcome. We focused on clinical safety as the key outcome, preferring to use the inclusion of all 'clinically important' diagnoses in the workup as the main variable of interest. In acute settings such as emergency rooms and primary care, where an incomplete and evolving clinical picture results in considerable diagnostic uncertainty at assessment, the ability to generate a focused and 'safe' workup is a more clinically relevant outcome, and one which accurately reflects the nature of decision making in this environment [53]. Consequently, we defined diagnostic errors of omission at assessment as the 'failure to consider all clinically important diagnoses (as judged by an expert panel working the same cases)'. This definition resulted in the 'correct' diagnosis, as well as other significant diagnoses, being included within the 'minimum' workup. Further, changes in test-ordering and treatment decisions were uniquely measured in this study as a more concrete marker of the impact of diagnostic decision support on the patient's clinical management; we were able to demonstrate an improvement in test-ordering in 1 in 10 system consultations, indicating that diagnostic DSS may strongly influence patient management, despite only offering diagnosis-related advice. Finally, the time expended during DSS consultation is an important aspect that has not been fully explored in previous studies. In our study, subjects spent a median of 6 minutes for clinical data entry (including typing in their unaided decisions), and a median of 1 minute to process the advice provided and make changes to their clinical decisions.

The research design employed in this study allowed us to confirm a number of observations previously reported, as well as to generate numerous unique ones. These findings relate to the operational consequences of providing diagnostic assistance in practice. In keeping with other DSS evaluations, different subject grades processed system advice in different ways, depending on their prior knowledge and clinical experience, leading to variable benefit. Since ISABEL merely offered diagnostic suggestions, and allowed the clinician to make the final decisions (acting as the 'learned intermediary') [54], in some cases, subjects ignored even important advice. In some other cases, they added irrelevant decisions or deleted important decisions after DSS consultation, leading to reduced net positive effect of the DDSS on decision making. For some subjects whose pre-DSS performance was high, a ceiling effect prevailed, and no further improvement could be demonstrated. These findings complement the results of our earlier system performance evaluation which solely focused on system accuracy and not on user interaction with DDSS. One of the main findings from this study was that consultants tended to generate shorter diagnostic workup lists containing the 'most likely' diagnoses, with a predilection to omit other 'important' diagnoses that might account for the patient's clinical features, resulting in a high incidence of DEO. Medical students generated long diagnostic workup lists, but missed many key diagnoses leading to a high DEO rate. Interestingly, all subject grades gained from the use of ISABEL in terms of a reduction in the number of DEO, although to varying degrees. Despite more DEOs occurring in cases considered to be routine in practice than in rare and difficult ones in the pre-DSS consultation phase, ISABEL advice seemed to mainly improve decision making for difficult cases, with a smaller effect on easy cases. The impact of DSS advice showed a decreasing level of beneficial effect from diagnostic to test-ordering to treatment decisions. Finally, although the time taken to process cases without DSS advice in this study compared favorably with the Friedman evaluation of QMR and ILIAD (6 min vs. 8 min), the time taken to generate a revised workup with DSS assistance was dramatically shorter (1 min vs. 22 min).

We propose a number of explanations for our findings. There is sufficient evidence to suggest that clinicians with more clinical experience resort to established pattern-recognition techniques and the use of heuristics while making diagnostic decisions [55]. While these shortcuts enable quick decision making in practice, and work successfully on most occasions, they involve a number of cognitive biases such as 'premature closure' and 'confirmation bias' that may lead to incomplete assessment on some occasions. On the other hand, medical students may not have developed adequate pattern-recognition techniques or acquired sufficient knowledge of heuristics to make sound diagnostic decisions. It may well be that grades at an intermediate level are able to process cases in an acute setting with a greater emphasis on clinical safety. This explanation may also account for the finding that subjects failed to include 'important' diagnoses during the assessment of easy cases. Recognition that a case was unusual may trigger a departure from the use of established pattern-recognition techniques and clinical shortcuts to a more considered cognitive assessment, leading to fewer DEO in these cases. We have shown that it is possible to reduce DEOs by the use of diagnostic reminders, including in easy cases, although subjects appeared to be more willing to revise their decisions for difficult cases on the basis of ISABEL suggestions. It is also possible that some subjects ignored relevant advice because the system's explanatory capacity was inadequate and did not allow subjects to sufficiently discriminate between the suggestions offered. User variability in summarizing cases may also explain why variable benefits were derived from ISABEL usage – subjects may have obtained different results depending on how they abstracted and entered clinical features. This user variability during clinical data entry has been demonstrated even with use of a controlled vocabulary in QMR [56]. We observed marked differences between users' search terms for the same textual case; however, diagnostic suggestions did not seem to vary noticeably. This observation could be partially explained by the enormous diversity associated with various natural language disease descriptions contained within the ISABEL database, as well as by the system's use of a thesaurus that converts medical slang into recognized medical terms.

The diminishing level of impact from diagnostic to test-ordering to treatment decisions may be a result of system design – ISABEL does not explicitly state which tests and treatments to perform for each of its diagnostic suggestions. This advice is usually embedded within the textual description of the disease provided to the user. Study and system design may both account for the differences in time taken to process the cases. In previous evaluations, subjects processed cases without using the DSS in the first instance; in a subsequent step, they used the DSS to enter clinical data, record their clinical decisions, and processed system advice to generate a second diagnostic hypothesis list. In our study, subjects processed the case and recorded their own clinical decisions while using the DSS for clinical data entry. The second stage of the procedure only involved processing ISABEL advice and modifying previous clinical decisions. As such, direct comparison between the studies can be made only by the total time involved per case (30 min vs. 7 min). This difference could be explained by features in the system's design that resulted in shorter times to enter clinical data and to easily process the advice provided.

The findings from this study have implications specifically for ISABEL as well as other diagnostic DSS design, evaluation and implementation. It is well recognized that the dynamic interaction between user and DSS plays a major role in their acceptance by physicians [57]. We feel that adoption of the ISABEL system during clinical assessment in real time is possible even with current computer infrastructure, providing an opportunity for reduction in DEO. Its integration into an EMR would allow further control on the quality of the clinical input data as well as provision of active decision support with minimum extra effort. Such an ISABEL interface has currently been developed and tested with four commercial EMRs [58]; this integration also facilitates iterative use of the system during the evolution of a patient's condition, leading to increasingly specific diagnostic advice. The reminder system model aims to enable clinicians to generate 'safe' diagnostic workups in busy environments at high risk for diagnostic errors. This model has been successfully used to alter physician behavior by reducing errors of omission in preventive care [59]. It is clear from recent studies that diagnostic errors occur in the emergency room for a number of reasons. Cognitive biases, of which 'premature closure' and faulty context generation are key examples, contribute significantly [60]. Use of a reminder system may minimize the impact of some of these cognitive biases. When combined with cognitive forcing strategies during decision making, DDSS may act as 'safety nets' to reduce the incidence of omission errors in practice [61]. Reminders to perform important tests and treatment steps may also allow a greater impact on patient outcome [62]. A Web-based system model in our study allowed users from disparate parts of the country to participate in this study without need for additional infrastructure or financial resources, an implementation model that would minimize the cost associated with deployment in practice. Finally, the role of DDSS in medical education and training needs formal evaluation. In our study, medical students gained significantly from the advice provided, suggesting that use of DDSS during specific diagnostic tasks (e.g. problem-based case exercises) might be a valuable adjunct to current educational strategies. Familiarity with DDSS will also predispose to greater adoption of computerized aids during future clinical decision making.

The limitations of this study stem mainly from its experimental design. The repeated measures design raises the possibility that some of the beneficial effects seen in the study are a result of subjects 'rethinking' the case, or the consequence of a reflective process [63]. Consequently, ISABEL's effects in practice could be related to the extra time taken by users in processing cases. We believe that any such effects are likely to be minimal since subjects did not actually process the cases twice during the study – a summary of the clinical features was generated by subjects when the case was displayed for the first time, and subjects could not review the cases while processing ISABEL suggestions in the next step. Subjects also spent negligible time between their first assessment of the cases and processing the diagnostic suggestions from the DSS. The repeated measures design provided the power to detect differences between users with minimal resources; a randomized design using study and control groups of subjects would have necessitated the involvement of over 200 subjects. The cases used in our study contained only basic clinical data gained at the time of acute assessment, and may have proved too concise or easy to process. However, this seems unlikely since subjects only took an average of 8 min to process even diagnostic conundrums prior to DSS use when 'expert systems' were tested. Our cases pertained to emergency assessments, making it difficult to generalize the results to other ambulatory settings. The ability to extract clinical features from textual cases may not accurately simulate a real patient encounter where missed data or 'red herrings' are quite common. The inherent complexity involved in patient assessment and summarizing clinical findings in words may lead to poorer performance of the ISABEL system in real life, since its diagnostic output depends on the quality of user input. As a corollary, some of our encouraging results may be explained by our choice of subjects: a few were already familiar with summarizing clinical features into the DSS. Subjects were not supervised during their case exercises since they may have performed differently under scrutiny, raising the prospect of a Hawthorne effect [64]. The use of a structured website to explicitly record clinical decisions may have invoked the check-list effect, as illustrated in the Leeds abdominal pain system study [65]. The check list effect might also be invoked during the process of summarizing clinical features for ISABEL input; this may have worked in conjunction with 'rethinking' to promote better decision making pre-ISABEL. We also measured decision making at a single point in time, making it difficult to assess the effects of iterative usage of the DSS on the same patient. Finally, our definition of diagnostic error aimed to identify inadequate diagnostic workup at initial assessment that might result in a poor patient outcome. We recognize the absence of an evidence-based link between omission errors and diagnostic adverse events in practice, although according to the Schiff model [53], it seems logical to assume that avoiding process errors will prevent actual errors at least in some instances. In the simulated setting, it was not possible to test whether inadequate diagnostic workup would directly lead to a diagnostic error and cause patient harm. Our planned clinical impact assessment in real life would help clarify many of the questions raised during this experimental study.

This experimental study demonstrates that diagnostic omission errors are common during the assessment of easy as well as difficult cases. The provision of patient- and context-specific diagnostic reminders has the potential to reduce these errors across all subject grades. Our study suggests a promising role for the use of future reminder-based DSS in the reduction of diagnostic error. An impact evaluation, utilizing a naturalistic design and conducted in real life clinical practice, is underway to verify the conclusions derived from this simulation.