This observational cohort study was conducted at 14 hospitals that were affiliated with BJC HealthCare and Washington University School of Medicine. These hospitals included both academic and community facilities, serving a diverse patient population across urban, suburban, and rural areas in Missouri and Southern Illinois.

All hospitals included in the study used the Epic EHR system. Epic’s Secure Chat, a text-based, EHR-integrated secure messaging platform, has been in use since 2019. Epic’s Secure Chat enables clinicians to send and receive text-based messages with other health care professionals who are part of the same institution directly from within the EHR interface [6]. For both the desktop and mobile versions of the Epic EHR, messages appear in conversational threads.

This study included all attending physicians, advanced practice providers (APPs; nurse practitioners and physician assistants), and trainee physicians (residents and fellows) who worked in inpatient settings during the 3-month period from February 1, 2023, to April 30, 2023.

An inpatient clinician’s workday (ie, a clinician-day) was included if they met the following criteria: (1) the clinician signed at least 1 patient note, (2) placed at least 1 order, and (3) sent or read at least 1 secure message. These criteria were used to ensure that the sample included active clinical workdays (ie, days in which the clinician was involved in direct patient care). This approach, consistent with prior research [22], reasonably excludes days when a clinician was not on service, as signing notes and placing orders are typically associated with direct patient care.

Clinician-days were excluded if they were night shifts, as night shifts have unique workflow dynamics, potentially affecting secure messaging behaviors [23,24]. Although night shift work is common, including such shifts could introduce heterogeneity in workflow patterns and bias the estimates. Night shifts were defined by the clinician’s first EHR login time occurring before 5 AM or after 5 PM (Figure S1 in Multimedia Appendix 1).

Additionally, at the clinician level, we excluded 45 clinicians who lacked demographic information (age and sex) from the analysis. For more information, see the flow diagram, with the inclusion and exclusion criteria, in Figure S2 in Multimedia Appendix 1.

In total, 2 primary data sources were used for this study: secure messaging metadata and EHR audit logs.

Secure messaging metadata were extracted from Epic’s Clarity database. For each message, we retrieved data including details about the conversation thread, message sender, message recipient, and timestamp that messages were sent and read. Secure messaging metadata and EHR-based audit logs were aggregated at the clinician-day level, defined as a 24-hour period within each calendar day.

EHR audit logs were retrieved from the ACCESS_LOG table in Epic’s Clarity database for the study period. These audit logs, which are required for compliance with Health Information Portability and Accountability Act, provide detailed trails of clinician activities within the EHR. EHR audit logs capture all click-based actions in the EHR that display or modify patient data, offering a comprehensive record of clinical work activities [25,26]. Audit logs have been extensively used in prior research to gain insights into clinician workload, workflows, and clinical behaviors [20,27].

Additionally, clinician-level metadata, including clinician role, clinical service locations (specific clinical settings in which they worked on a particular day), and demographic details were also extracted for analysis.

Finally, clinicians, especially trainees, may have worked in multiple settings over the study period. To account for this, we used the clinical service assignment for each clinician-day based on their EHR login information. When logging into the EHR, clinicians select a “context” for their login, providing them with a customized EHR interface aligned with that clinical service. The most frequent login context for each clinician-day was used as a proxy for their clinical service assignment.

To identify conversational multitasking, we first sought to determine the occurrence of concurrent secure messaging conversations during a clinician-day. Towards this end, we defined a conversation as active from the first to the last timestamp of the clinician’s secure message activity within that conversation for each day [28]. For each minute of the clinician-day, we then computed the number of active secure messaging conversations that the clinician was engaged in. For example, Figure 1 illustrates the messaging timeline for a clinician engaged in 3 conversation threads from 9:30 AM to 10:30 AM. Conversation 1 started at 8:30 AM and ended at 11:00 AM; Conversation 2 started at 8:40 AM and ended at 10:30 AM; Conversation 3 started at 9:30 AM and ended at 11:30 AM. The number of concurrent secure messaging conversations at each 1-minute interval was calculated as follows: 8:30 AM to 8:40 AM (0 concurrent); 8:40 AM to 9:30 AM (2 concurrent); 9:30 AM to 10:30 AM (3 concurrent); 10:30 AM to 11:00 AM (2 concurrent); and 11:00 AM to 11:30 AM (0 concurrent). Sensitivity analyses were also performed using 2- and 5-minute intervals to assess the robustness of the findings (Tables S3-S6 in Multimedia Appendix 1).

All data processing and visualization were performed using Python 3.9.10, Pandas 2.2.2, and Matplotlib 3.9.1 [29-31].

The exposure of interest was the level of conversational multitasking, defined as the maximum number of concurrent secure messaging conversations a clinician engaged in during a workday. We categorized the maximum number of concurrent secure messaging conversations for a clinician-day into 4 groups: no concurrent secure messaging conversations, 2 concurrent secure messaging conversations, 3 concurrent secure messaging conversations, and ≥4 concurrent secure messaging conversations.

We conceptualized the maximum number of concurrent conversations as a categorical variable, as the integer values were highly skewed. Moreover, we did not expect the relationship between the number of concurrent conversations and the outcomes of interest to remain constant for each 1-unit increase in concurrency. These specific category boundaries were selected to ensure a relatively balanced distribution of clinician-days across groups, which is important for obtaining a representative picture of the entire population under study and for ensuring unbiased and efficient estimates [32,33]. For instance, as shown in Figure 1, this clinician’s highest number of overlapping secure messaging conversations was 3, so they would be classified in the 3 concurrent secure messaging conversations group. Conversely, if none of the clinician’s secure messaging conversations overlapped during the day, the clinician-day would be categorized in the 0 concurrent secure messaging conversations group.

We used 2 outcome measures: EHR time and patient switching. EHR time, measured in minutes, was calculated from the audit log data by summing the time intervals between successive audit log actions, excluding intervals exceeding 5 minutes, which were considered periods of inactivity. This method for estimating overall EHR time is commonly used as an indicator of clinician workload [34-36]. Prolonged EHR usage has been linked to adverse wellness outcomes, including burnout [34,37,38].

Patient switching was also derived from audit log activities and was computed as the frequency of transitions between patient charts during an uninterrupted EHR session (also referred to as attention switching) [22]. A transition between audit log events in one patient’s chart to another patient’s chart was recorded as a patient switch, unless the interval between the events was greater than 5 minutes. Cognitive science research identifies attention switching as a cognitive burden, which has been associated with an increased risk of errors [36,39].

For each clinician-day, we calculated the total time spent in active secure messaging conversations, representing the cumulative minutes engaged in any messaging. We then determined the percentage of time spent in concurrent secure messaging conversations by dividing the minutes with concurrent secure messaging conversations by the total messaging time.

Other covariates included daily secure messaging volume, patient load, clinician role, clinical service, age, and sex.

Descriptive statistics were calculated as medians with IQRs or as frequencies with percentages. In total, 2 separate linear mixed effects models were used to investigate the relationship between the maximum number of concurrent secure messaging conversations and both total EHR time and patient switching. Models included the maximum number of concurrent secure messaging conversations as the primary exposure of interest and were adjusted for daily messaging volume, patient load, percentage of time with concurrent secure messaging conversations, clinician role, age, and sex. Models included random intercepts for clinicians to account for repeated observations within clinicians and for clinical service to account for similarities in work patterns within these settings. Because each clinical service was nested within a single institution, this helped to account for institutional-level variation in clinical work patterns. Linear mixed effects models were the preferred statistical approach, as opposed to alternatives such as generalized estimating equations, as it was important to account for clustering among both clinicians and the clinical service where they worked to ascertain the effect estimates after accounting for these variances.

Effect estimates and 95% CIs were calculated to evaluate the expected association between the maximum number of active secure messaging conversations on a clinician-day and the outcome variables. The interaction between the maximum number of concurrent secure messaging conversations on a clinician-day and clinician role was also assessed in a separate model. Based on the significance of the interaction term, 3 separate models, stratified by clinician role, were subsequently run for each outcome variable to explore these associations within specific clinician groups [40]. All statistical analyses were conducted using R (version 4.3.3; R Foundation for Statistical Computing) and RStudio (version 2023.12.1; Posit, PBC) [41,42].

This study included 50,027 clinician-days, covering >80 million audit log actions and 3,556,562 secure messages by 3,232 unique clinicians (Figure S2 in Multimedia Appendix 1). The median clinician age was 37 years (IQR 32‐46), and 1798 (56%) were female. Clinician roles were distributed as follows: 54% (1759) were attending physicians, 27% (857) were APPs, and 19% (616) were trainee physicians. The median patient load per day was 7 (IQR 4‐11). Clinicians spent a median of 307 minutes per day on the EHR (IQR 204‐413), with a median of 107 patient switches per day (IQR 60‐176). Detailed descriptive characteristics by clinician role are provided in Table 1.

For secure messaging activities, the median volume of secure messages sent and received per day was 37 (IQR 13‐88), with a median of 203 (IQR 42‐430) minutes per day spent on messaging. Concurrent secure messaging conversations occurred on 28,151 (56%) clinician-days, with a median of 34% (IQR 11%‐62%) of messaging time spent in concurrent secure messaging conversations. Further details on secure messaging time across different concurrency levels are shown in Table S7 in Multimedia Appendix 1.

In multivariable analysis, compared to clinician-days with no concurrent secure messaging conversations, clinician-days with a maximum of 2, 3, and 4 or more concurrent secure messaging conversations were associated with an increase in the total EHR time of 20.3 (95% CI 18.2‐22.4), 38.0 (95% CI 34.9‐41.1), and 54.8 (95% CI 50.6‐58.9) minutes, respectively (all P<.001).

The interaction between the maximum number of concurrent secure messaging conversations and clinician roles was statistically significant (P<.001). Consequently, we conducted a stratified analysis based on the clinician role (attending physician, APP, and trainee physician). Results were maintained within all 3 subgroups, but the associations were strongest among trainee physicians and weakest among APPs (Table 2).

Specifically, for trainee physicians, clinician-days with a maximum of 2 concurrent secure messaging conversations were associated with an increase in EHR time of 30.6 minutes (95% CI 25.6‐35.5); clinician-days with a maximum of 3 concurrent secure messaging conversations were associated with an increase of 58.3 minutes (95% CI 51.4‐65.3); and clinician-days with 4 or more concurrent secure messaging conversations were associated with an increase of 82.3 minutes (95% CI 73.2‐91.5), all compared to clinician-days with no concurrent secure messaging conversations.

In multivariable analysis, compared to having no concurrent secure messaging conversations on a clinician-day, clinician-days with a maximum of 2, 3, and 4 or more concurrent secure messaging conversations were associated with 14.5 (95% CI 11.3‐17.7), 26.7 (95% CI 21.9‐31.5), and 41.6 (95% CI 35.2‐48.1) additional patient switches, respectively.

As with EHR time, there was a statistically significant interaction between the maximum number of active secure messaging conversations and clinician roles for patient switching (P<.001). Again, results were maintained in all 3 subgroups, and the strongest association was among trainee physicians (Table 3).

For trainee physicians, clinician-days with a maximum of 2 concurrent secure messaging conversations were associated with 23.3 (95% CI 19.2‐27.3) additional patient switches; clinician-days with 3 concurrent secure messaging conversations were associated with 43.4 (95% CI 37.7‐49.1) additional patient switches; and clinician-days with 4 or more concurrent secure messaging conversations were associated with 61.8 (95% CI 54.3‐69.3) additional patient switches, all compared to clinician-days with no concurrent secure messaging conversations.

The concurrency among actively engaged secure messaging conversations was captured at a per-minute level (Tables S1 and S2 in Multimedia Appendix 1). We performed sensitivity analysis with additional time windows (2 and 5 min), and there were no meaningful changes in the effect sizes for both outcomes (Tables S3-S6 in Multimedia Appendix 1).

In this large-scale cohort study with >3000 clinicians across more than 50,000 clinician-days, encompassing >3 million secure messages and >80 million audit log actions, we found that increased conversational multitasking was associated with increased EHR time and more patient switches in a dose-dependent manner. This highlights the potential association between conversational multitasking and higher clinical workload and cognitive burden. Notably, higher levels of conversational multitasking were associated with progressively higher clinical workload and cognitive burden.

A recent, related study, on which the current study was based, found an association between secure messaging volume and both clinician workload and attention switching [22]. The current study examined the impact of inpatient clinicians engaging in concurrent secure messaging conversations and emphasizes the role of conversational multitasking on clinician burden: even after adjusting for messaging volume, higher levels of conversational multitasking were associated with considerably larger increases in both EHR time and patient switches.

This increased burden potentially arises from interleaved multitasking associated with managing active, ongoing messages along with other clinical and EHR-based tasks. Prior research has shown that when users were prepared for multiple ongoing tasks (ie, multitasking), their task performance became slower, even if the secondary task was not directly in their view [14]. In other words, the simultaneity of active secure messaging conversations increases the user’s cognitive burden, as the user’s cognitive resources (ie, working memory) are overloaded because of the superimposition of the multiple ongoing messages [43].

The interleaving of work activities has been shown to affect clinician behaviors in a number of settings [15,17,19]. In contrast to most of which included observational techniques with small sample sizes, the current study (with >3000 clinicians) highlights the potential impact of interleaving of clinician work activities on downstream outcomes (ie, clinical work and cognitive burden). Moreover, multitasking using secure messaging may also have patient safety implications. A recent study showed that increased volume of secure messaging may be associated with a corresponding increase in the likelihood of wrong-patient ordering errors [44]. As part of future work, we will investigate the association between conversational multitasking and errors, care delays, or other adverse patient safety events.

Finally, it must be stated that secure messaging has many potential advantages [45]; it is an asynchronous communication mode that allows for the ability to prioritize messages, thus potentially streamlining clinical workflows [46,47]. Prior studies have also found that clinicians perceive secure messaging as being easier to use and better integrated with workflow, potentially improving communication efficiency [4,9,10,47]. However, given the rapid proliferation of the use of digital tools in medicine, it is important to consider their impact on clinicians and their work activities. Increasing use of digital tools has increased clinician work on the EHR, work outside of work, and more concerningly, potentially reduced bedside time with patients and increased clinician burnout [48]. As such, with secure messaging, it is important to carefully consider policies and interventions that can lead towards more efficient use of these digital tools [49,50].

This study has several limitations. First, this study was conducted in the inpatient settings of a single integrated health care system, which may limit its generalizability; however, it did include 14 academic and community hospitals, with a large sample of clinicians (>3000) and their workdays (>50,000). Second, communication needs may also differ based on specific clinical settings and associated workloads (eg, an intensive care unit compared to a hospital ward). We adjusted for this by controlling for the clinical setting and accounting for each clinician’s patient load, but we cannot exclude the possibility of residual confounding. Third, we assumed concurrent conversations contributed to clinician cognitive burden throughout the entire time they were open, and thus we allowed conversations to contribute to concurrency for the entire duration that they were open. However, many conversational threads evolved over extended periods (minutes to hours); as such, it is possible that clinicians may have forgotten about some conversational threads over time, so our conceptualization of concurrency may potentially overestimate the impact of these conversations on cognitive burden. Similarly, the observed association between conversational multitasking, EHR time, and patient switching may be confounded by higher care coordination needs on that workday, as we were unable to directly adjust for patient complexity. However, we could not adjust for patient complexity, primarily because a considerable percentage of the messages (~30%) did not include patient identifiers (eg, a medical record number), making it difficult to link secure messaging conversations to specific patients. Fourth, 45 clinicians were not included because of the lack of their demographic information, potentially introducing selection bias; however, the number of excluded clinicians was small relative to the overall cohort (approximately 1%, 45 out of 3979). Fifth, measurement of the EHR time may be imprecise because it was derived from EHR-based audit logs. However, recent studies have validated that such audit log-based measures of EHR time are closely correlated with observed time spent on the EHR [20,51,52]. Finally, secure messaging was not the only form of communication available; other forms of communication included face-to-face, telephone, and pagers. As such, the overall burden of communication could not be measured. However, as has been shown in recent research, new channels of communication lead to polychronic communication—use of multiple channels for communication (eg, secure messaging and phone)—increasing the overall burden [53].