In the last two decades, the percentage of physicians using an electronic health record (EHR) in their practice has risen from around 10% to over 80%, driven in varying degrees by federal initiatives, intra- and internetwork pressure, and perceived potential advantages. This increase represents a broad shift in practice patterns that have introduced new challenges into the process of delivering quality medical care. Though EHRs come with many benefits such as improved compliance with clinical guidelines, substantial decreases in medication errors, the ability to improve care through simulation, and increased cost-savings at an institutional level, the development and implementation of the EHR has far outpaced investigation into their downsides [1-5]. It is clear that the amount of documentation required per patient has increased substantially with the advent and implementation of EHR systems [6]. Interface design, usability, and interoperability issues have plagued EHRs since their initial iteration in the 1970s; increased documentation and input requirements correlate with increased error rates, longer hospital stays, less job satisfaction, and physician burnout [7-10]. Tools for understanding the usability patterns of EHRs are sorely needed in order to better inform system design and provider training.

The user experience of any type of electronic system such as an EHR depends entirely around the ways in which the user navigates the system to produce desired outcomes; an eminently capable program will be useless if users cannot utilize its capabilities. Appropriately then, interface navigability is a consideration at the forefront of electronic system design. Of course, humans did not evolve with electronic system navigation as an active evolutionary pressure; instead, the human brain is optimized for spatial navigation. There are two recognized spatial navigation strategies: (1) local landmark usage and (2) distant cues (eg, cardinal directions) utilization [11]. All neurotypical humans are capable of utilizing either strategy when required by a given situation; however, individuals have a reliable tendency to default to one strategy over the other, and sex is a strong predictor of one’s default strategy [11-15]. In these spatial tasks, females tend to default to using local landmarks while males default to utilizing distal cues. Although social expectations can have an effect on task performance, the phenomenon of default strategy selection appears to be strongly influenced by genetic and hormonal factors rather than socialization [16-19].

Analogous sex differences in navigational strategy have also been observed in the varied contexts of navigating electronic screens. Females reported higher utilization of situation-specific navigational cues during single-site hypertextual (ie, website) navigation and more frequently used landmarks to guide others verbally when acting as instructors [20]. Previous work showed that performance on text-based database querying tasks (eg, Google, PubMed, etc) improves with increased content knowledge across both sexes; however, no matter the level of content knowledge, elementary-school–aged females spent more time reading documents, formulated fewer overall queries, and clicked on fewer links per minute than their male counterparts, although none of these measures correlated with overall task performance [21]. Constrained to a search task in an area of low-domain knowledge, middle-school–aged males and females showed no differences in performance in nonelectronic library searches but did show a difference in task performance when using the web, a finding which has been replicated [22,23]. Finally, differences in gaze patterns during the higher-level analytical tasks of C-program debugging correlated with task performance [24,25].

Only a handful of studies have investigated eye tracking with regard to EHR usage with the majority focused on EHR use as opposed to characterizing user characteristics [26-28]. Previously, our group used eye tracking to characterize general features of gaze in hospital physicians and found that there were two distinct workflows: while all participating physicians spent far more time writing their notes than perusing the chart overall, one group of physicians perused the chart at length before opening their notes while the other started their notes immediately. This difference in workflow was associated with physician sex in that females physicians were more likely to peruse the chart before starting their notes [29]. Though no sex-related performance differences in medical diagnostic reasoning have been reported thus far, these results showed that differences in EHR navigation do exist between female and male physicians. While the impact of sex differences on EHR navigation remains unknown, what is apparent is that the increase in EHR use nationally has correlated with a dramatic increase in physician burnout that is increasingly being recognized as disproportionately affecting female physicians [9,10,30].

Simulation affords a powerful tool to better understand the role of sex in EHR navigation by removing the variability inherent in analyzing deidentified patient charts and replacing it with a standardized, validated case that is the same for all participants. We have previously reported the use of high-fidelity EHR-based simulation to understand error recognition in simulated intensive care unit (ICU) charts [28,31-33]. In these preliminary studies, we incorporated eye and screen tracking to determine whether gaze metrics could be used as surrogates for EHR performance. We demonstrated that a number of eye-tracking metrics correlated with the recognition of embedded safety items within the chart for the entire cohort. The goals of this study were twofold: (1) to expand upon the initial data set in order to analyze whether sex was associated with differential navigation patterns during simulated EHR exercises, and (2) to determine the impact, if any, this difference had on task performance.

This study was approved by the Institutional Review Board at Oregon Health & Science University. The study was deemed minimal risk, and formal informed consent was not required. However, all participants were provided with an information sheet on our research protocol.

Physicians’ eye movements were recorded with a Tobii X1 monitor–mounted Light Eye Tracker (Tobii Systems) while completing a previously validated EHR-based simulation exercise in our EHR environment (EpicCare, Epic Systems Inc) [4,33]. This simulated instance imported all end-user settings, preferences, and customized screens so that participants were able to use their own personalized data gathering tools. We created 2 simulated ICU patients, each admitted to the ICU for 5 days. We attempted to make the simulated patient scenarios as robust as possible, with hourly vital signs, lab results, and nursing, as well as resident and attending notes. Residents and fellows rotating through the medical ICU were eligible for enrollment, and thus all had experience with the use of the EHR in the context of the exercise they were being asked to complete. Participants were provided a written “signout” and then given 10 minutes to review data on one of the simulation cases, with instructions to review the EHR as if they were assuming care for the patient and would be presenting them on rounds, along with any potential changes in management that would be required for their care. At the end of 10 minutes, participants then presented the care plan for the patient to a member of the study team. The participants’ performance was assessed by the number of safety issues they verbalized during this presentation, followed by a debriefing of the safety issues associated with the case as previously described [4,31,32]. The physical nature of the computer station used for the testing was standardized by fixing monitor, desk, and chair height, as well as tilt and relative positioning of the monitor at 65 cm from the participant. Calibration was performed for each subject using a 1-minute–long, 9-point calibration algorithm provided by the manufacturer.

Once complete, all videos were captured and analyzed with the vendor’s eye-tracking software (Tobii Studio, Tobii Systems). Fixations were defined as a period (>60 milliseconds) in which providers’ eyes tracked without eye velocity exceeding 30 degrees/second. Saccades (rapid scanning eye movements) were defined as periods with gaze velocity exceeding 30 degrees/second, these definitions being adapted from current standards within the literature [34]. Static eye tracking on points for <60 milliseconds were termed “microfixations,” as they did not meet broadly acceptable standards for the fixations referenced above. These were present primarily while participants were retrospectively observed to be reading blocks of full sentences. Raw tracking data was analyzed using custom Excel macros designed to calculate more complex variables (eg, screen distance between fixation points, fixation duration, etc). Variable distribution was analyzed in GraphPad (GraphPad Software Inc). Many of the recorded variables were definitionally bounded on the low end producing one-tailed variable distributions. Given these distributions, appropriate nonparametric tests were used to assess for differences between group means.

Krejtz et al [35] introduced coefficient Κ, seeking to construct a single, trackable indicator of fluctuation between ambient or focal eye movements. This metric was designed to track gaze patterns across visual stimuli (eg, art, visual stimulation tasks, etc), and their reported results are promising; however, it was not suitable for answering the question posed in our work, given that we sought to differentiate between two active self-driven data extraction strategies (“scanning” vs “staring”), which requires a variable built to purpose [35]. We thus developed a novel variable—the Gaze Index (GI)—to denote the ratio of time spent in long fixation (“staring”) to time spent in saccadic motion (“scanning”). The GI represents the proportion of time each participant spent in fixation versus saccade. For example, a GI of 0.11 would indicate that a participant spent approximately 9 times more of their screen time in saccade-microfixation-saccade-etc (a “reading”/“scanning” pattern) than in full fixation (“staring”). It inherently normalizes for the time each participant spent viewing the screen, making the GI a superior representation of participant behavior when compared to the raw ratio of fixations to saccades. Normalization to screen time is particularly important given the top-bounded nature of saccadic movement compared with the unbounded upper limit of fixation periods, which could easily skew results based on the finite nature of this time-constrained task.

For analysis, we correlated individual usability measures to performance on the simulation. Differences in performance between groups with specific usability characteristics were compared using a Mann-Whitney U Test based on the size of the data set and the nonparametric nature of the data. Normality was assessed via Kolmogorov-Smirnov. All analyses were performed using GraphPad Prism (GraphPad Software Inc), and a P value of <.05 was considered statistically significant.

The data set comprised the eye-tracking recordings of 93 physicians (female: n=46, male: n=47). Of these, 39 (42%) were interns; the remaining were in their PGY (postgraduate year) 2 year and above. None of the recordings were excluded from analysis. Self-supplied specialty identifications were similar between groups (Table 1).

Overall, females spent less task time looking at the screen than males (44.7% vs 55.1%; P=.007, U=767) (Figure 1).

While looking at the screen, the mean fixation (“stare”) duration of all participants was 212 milliseconds, which is in line with previously reported mean fixation durations during text-reading tasks [36]. No difference was found in the number of fixations per minute between females and males (mean 191.3, SD 5.1 vs mean 189.7, SD 3.9) (Figure 2A); however, the mean fixation duration was shorter for females than males (mean 197.6, SD 7.2 vs mean 226, SD 6.6 milliseconds; P=.005, U=740) (Figure 2B). Finally, there was a trend toward an increase in the mean distance traveled between fixations among females (mean 195.7, SD 5.1 pixels vs mean 184.6, SD 4.5 pixels; P=.07, U=892) (Figure 2C).

We next analyzed saccadic (“scanning”) eye movements. Females traveled more distance per minute of screen time (35,026, SD 1065 pixels per minute vs 33,542, SD 943 pixels per minute; P=.03, U=839) (Figure 3A) and had more saccades per minute of screen time (347, SD 18 vs 288, SD 16; P=.02, U=817) (Figure 3B). Furthermore, their periods of saccadic movement were of longer duration (63.1, SD 2.4 milliseconds vs 56.4, SD 1.8 milliseconds; P=.03, U=836) (Figure 3C). In total, female physicians scanned more often and for longer durations than their male counterparts.

In order to investigate the relative time spent between fixation and saccadic movement, we calculated GI values for all subjects. The GI had a wide frequency distribution across all participants (0.11-7.02). The mean GI of all participants was approximately a 3:1 ratio, indicating that the participants, on average, spent 3 times as much of their screen time staring as they did scanning: 46 participants had a GI value >3.0, while 47 participants had a GI value <3.0 (Figure 4).

We termed those users with a GI>3 “starers” and those with GI<3 “scanners” to qualify their tendency toward either eye movement relative to the average GI of 3. Interestingly, females were more likely than males to have a GI<3.0 (2.4 vs 3.4; P=.004, U=739) (Figure 5A). When looking at GI extremes, males were more likely to have a GI>5 (odds ratio [OR] 4.51 vs OR 0.22) while females were more likely to have a GI<1 (OR 1.58 vs OR 0.63) (Figure 5B). No differences between specialties were detected either in pooled analysis or in analysis within sex groups (not shown).

Lastly, we sought to determine whether differences in navigation patterns predicted performance in the simulation. Overall, there were no appreciable differences between females and males in task performance measured by percentage of existing safety items identified (39.9, SD 16% vs 38.6, SD 19%; P=.97) (Figure 6). Similarly, when users were grouped by their differences in GI rather than by sex (GI<1 vs GI>5; GI<3 vs GI>3) there were no detectable differences in task performance (not shown).