We drew on a sample of older adult patients aged 65 years and older who received care at a large academic health system between October 1, 2021, and September 30, 2022, as described in previous work [43,44]. Our sample comprised 16,023 unique patients, all of whom sent at least 1 secure “medical advice request” message during the year via a patient portal (MyChart). A medical advice request is one of several options used to categorize patient messages in this patient portal; this message type was selected because it is the most open-ended type of patient message and most likely to be directed to clinicians. Other patient message types (eg, “appointment scheduling request”) are likely to be routed to other staff (such as front desk staff) and were not included in our sample.

Patient demographic characteristics drawn from the electronic medical record (EMR) included patient age in years, sex (male or female), race (White, Black, Asian or Pacific Islander, or other: any other race), ethnicity (Hispanic or non-Hispanic), marital status (married or not married, divorced or legally separated, single, or widowed), and whether the patient identified English as their preferred language.

Because our EMR data did not include measures that reflect socioeconomic status, we used patient addresses and 9-digit zip codes to calculate an Area Deprivation Index (ADI) score for each patient, as a proxy measure of socioeconomic disadvantage. An ADI score is composed of 17 US Census–based measures that collectively characterize the socioeconomic disadvantage of a neighborhood [45,46]. ADI scores range from 1 to 100, with higher scores indicating a higher level of disadvantage in that neighborhood.

We calculated the Charlson Comorbidity Index (CCI) for each patient from International Statistical Classification of Diseases, Tenth Revision (ICD-10) codes to measure overall disease burden, with higher scores indicating greater comorbidity [45,47]. Finally, we identified the presence of a diagnosis of 3 specific health conditions: cancer, dementia, and diabetes. We focused on diabetes because of the robust literature describing patient portal and secure messaging use for this condition [17,48-50]. Dementia and cancer were included because of the complexity of care involved in these conditions and the relevance of care coordination, which secure messaging can facilitate [15].

We constructed 5 measures reflecting secure messaging intensity from time-stamped patient portal and EMR data. We conducted outlier analysis using the IQR method [51] to identify “high-intensity” message senders according to each measure. For each measure, the upper threshold was determined as quartile 3+1.5×IQR. Any values equaling or exceeding this upper bound were classified as upper outliers and thus “high-intensity” (Multimedia Appendix 1). Detailed descriptions and rationale for each measure are included in Multimedia Appendix 2 [10,52-54]. We briefly describe each of the 5 measures here.

First, the volume measure identifies high-intensity message senders based on the aggregate number of messages sent during the observation period. Any patient identified as an outlier by the IQR method is referred to as a “high-volume” message sender. This serves as our primary measure of high-intensity message sending and defines our main group of interest for analysis.

Second, the messaging episodes measure captures “episodes” of concentrated messaging. To conservatively allow time for responses, messages sent less than 10 days apart were considered part of the same messaging episode. Within each episode, we count the number of messages and identify “high-intensity episodes” using the IQR method.

Third, the length (character count) measure calculates the number of characters per message to identify messages that had a disproportionately high number of characters using the IQR method.

Fourth, the ratio of messaging to portal logins (message-to-portal ratio) measure is operationalized as the ratio of the number of secure messages to the number of patient portal logins during the observation period and reflects disproportionate messaging relative to portal use.

Fifth, the ratio of messages to encounters (message-to-encounter ratio) measure reflects a ratio of the number of secure messages to the number of in-person encounters during the year.

For each measure, we categorized patients as “high-intensity” or “other.” Beyond identifying high-intensity senders for the messaging episode measure (those who had at least 1 episode with a “high” number of messages) and the length measure (those who sent at least 1 disproportionately “long” message), we also report the number of messaging episodes meeting the criteria for “high” messaging volume by patient and the number of messages meeting the criteria for being “long,” by patient. The measures expressed as ratios are reported as continuous measures. We also separately report the number of portal logins and encounters.

This study was approved by the Johns Hopkins School of Medicine Institutional Review Board (IRB00303190). All data were housed and analyzed within a secure, access-controlled computing platform provided by Johns Hopkins University.

The sample consisted of 16,023 unique patients who sent a total of 199,981 messages over the 12-month observation period (Table 1). The average number of messages sent per patient was 12.5 (SD 20.4), with a median of 6 (IQR 12) messages (Multimedia Appendix 1). Patients had an average age of 73.4 (SD 7.0) years and were predominantly White (n=11,714, 73.1%), non-Hispanic (n=15,585, 95.4%), and married (n=10,528, 65.7%), and predominantly had a preferred language of English (n=15,719, 98.1%; Table 1). The average ADI among patients was 65.3 (SD 20.1). The average comorbidity score (CCI) was 1.9 (SD 1.7). In total, 26.1% (n=4174) of patients in our sample had diagnosed diabetes, 23.9% (n=3819) had diagnosed cancer, and 6.2% (n=996) had diagnosed dementia.

High-volume message senders were defined as patients who sent 33 or more messages during the 12-month observation period (Multimedia Appendix 1). High-volume message senders were more likely to be White (1044/1298, 80.4% vs 10,670/14,725, 72.5%) and older than other message senders (mean age 73.9, SD 6.9 years vs 73.4, SD 7.1 years; P=.006) but were otherwise similar with respect to sex, marital status, and ADI. High-volume message senders had a greater mean comorbidity score (2.6, SD 1.9 vs 1.8, SD 1.7; P<.001) and were more likely to have been diagnosed with cancer (465/1298, 35.8% vs 3354/14,725, 22.8%; P<.001) and dementia (107/1298, 8.3% vs 889/14,725, 6.1%; P<.002) than other message senders.

In adjusted analyses, older age (adjusted odds ratio [aOR] 0.99, 95% CI 0.98‐1.00), female sex (aOR 0.85, 95% CI 0.75‐0.96), and Black (aOR 0.45, 95% CI 0.37‐0.53) or Asian (aOR 0.29, 95% CI 0.16‐0.53) race (compared with White) and having a diagnosis of diabetes (aOR 0.63, 95% CI 0.54‐0.74) were associated with lower odds of being a high-volume message sender (Table 2). Higher comorbidity burden (aOR 1.33, 95% CI 1.28‐1.38) and a diagnosis of cancer (aOR 1.30, 95% CI 1.14‐1.48) were associated with greater odds of high-volume messaging. Ethnicity, marital status, dementia diagnosis, and ADI were not associated with high-volume messaging.

High-volume message senders comprised a small proportion of total message senders (1298/16,023, 8.1%; Figure 1) but sent 39.7% (79,475/199,981) of all messages. High-volume message senders on average sent 61.2 (SD 43.6) messages per year, nearly 8 times more than other senders, who averaged 8.2 (SD 7.5) messages annually (Table 3).

Nearly 9 in 10 (1160/1298, 89.4%) high-volume message senders experienced at least 1 high-intensity messaging episode (8.5 or more messages per episode), as compared with just 6% (883/14,725) of other senders (P<.001; Table 3). The number of high-intensity messaging episodes was significantly greater among high-volume message senders (mean of 2.4, SD 1.9 vs 0.1, SD 0.3; P<.001).

This analysis identified a total of 11,186 long messages (messages equal to or over 787.5 characters), representing approximately 5.6% of the total messages. High-volume senders were responsible for more than half of long messages (Figure 1). About 68% (872/1298) of high-volume message senders authored at least 1 long message, compared with 19.6% (2,883/14,725) of other senders (P<.001; Table 3). On average, high-volume message senders authored 3.8 (SD 13.0) long messages; in contrast, other senders on average sent less than 1 long message (0.3, SD 1.0).

High-volume secure message senders exhibited significantly greater overall patient portal use (Table 3). They had a higher message-to-portal login ratio (0.52 vs 0.15; P<.001) and sent more messages relative to health care encounters (message-to-encounter ratio 3.3 vs 1.3; P<.001) compared with other senders.

All but 2 of the 5 measures were weakly correlated (Pearson correlation coefficients ranging from 0.08 to 0.33; Table 4), indicating that the 5 measures captured distinct aspects of messaging behavior. Secure messaging volume and high-intensity messaging episode measures were most highly correlated (Pearson correlation coefficient of 0.68).

This observational study presented a profile of high-volume secure message senders across sociodemographic characteristics, health status, and messaging behaviors. We found that high-volume message senders comprised about 8% of message senders but accounted for nearly 40% of message volume. High-volume message senders were more often White and had a higher comorbidity burden. To more fully capture messaging behavior beyond aggregate volume, we developed 4 additional measures of messaging intensity. These revealed that high-volume message senders, compared with all other senders, were more likely to engage in sustained messaging, to send longer messages, to use the patient portal for functions beyond messaging, and to send more messages relative to their in-person health care encounters.

Our finding that all but 2 of the 5 measures of messaging intensity were weakly correlated underscores that patient-initiated messaging behaviors are nuanced. While a simple measure of aggregate messaging volume effectively captures patients who rely on messaging as a key form of communication with their clinicians, additional meaning can be derived from more granular measures of messaging behavior. Focusing on patients who meet multiple high-intensity messaging criteria could help resource-limited health systems identify patients with high communication needs who may benefit from additional clinical support. For example, patients who send long messages shortly after clinical encounters may benefit from a previsit agenda-setting tool to ensure that key concerns are addressed during the visit [49].

Our analysis found that high-volume senders sent, on average, 3 times as many long messages as all other message senders. Health systems and EMR vendors could also consider enhancing patient portal functionality to help clinicians more easily respond to lengthy messages. For example, clinicians may struggle to efficiently extract key details from long free-text messages [16,56]. A structured message template could help patients provide key information in an optimized format. Emerging AI-based approaches may further support message management by assisting with triage and prioritization of incoming messages based on urgency [57]. Some health systems also use AI to generate initial draft responses to patient messages, with drafts being reviewed and edited by clinicians before they are sent [58-60]. AI could be used to draft responses for only messages with a lower character count, protecting clinician time for responses to longer messages, which may be more complex and require greater clinical judgment.

Our finding that high-volume message senders had higher comorbidity burden (particularly with respect to cancer) and more extensive in-person health care use reinforces that high messaging volume may reflect greater health care needs. This highlights that efforts to manage secure messaging should focus not only on improving messaging efficiency but also on expanding health system capacity to meet the communication needs of patients with complex illnesses. A recent systematic review found that patients with chronic conditions find patient portal–based educational materials useful [61]. Health systems could embed curated educational materials within the patient portal to support patients with complex health conditions, providing information and referrals to community resources [62].

Current billing codes for clinician responses to patient messages are based solely on the amount of time spent replying to a message, without accounting for the clinical complexity or cognitive demands involved [22,23,32]. Our finding that a substantial proportion of messages come from patients with serious illnesses suggests that time-based billing codes may undervalue the clinical expertise and judgment required for high-quality messaging communication [23] when responding to messages from patients with complex care needs. Our findings may help inform refinements to billing models to more accurately reflect the level of effort and knowledge involved in supporting patients with complex care needs via secure messaging.

Importantly, input from clinicians, care teams, and patients is essential for developing effective strategies [63-65]. This should include understanding of which types of messages present the largest cognitive load or require the most effort to respond to, in which cases messaging is most helpful to patients, and identifying workflows and team structures that support optimal message handling. Costs of implementation and oversight requirements would need to be carefully considered with the implementation of any AI-based features. Ultimately, a multipronged approach to managing secure messaging volume, informed by patient needs, illness types, and communication preferences, can enhance both patient support and ease message-related burden on clinicians.

This study has several limitations. Our sample was constrained to patients receiving care from a single academic health system in one geographic region, limiting the generalizability of the findings. We did not capture health care that patients may have received elsewhere; patients may have sent messages to clinicians outside of our health system, and we may have underestimated their true messaging habits. We did not examine clinician factors that may influence patient messaging intensity, such as clinician response speed, response quality, or clinician encouragement to use secure messaging. Finally, as our analysis relied primarily on patient characteristics and information available within the EMR, we were unable to speak to external factors that influence messaging behaviors, such as digital literacy, social support, or preferences for other modes of communication. While our findings are limited to the inferences possible based on a single health system’s EMR, we present measures that can be readily replicated by other health systems to evaluate their patients’ unique messaging profiles and needs.

Our study is the first to compare multiple approaches to defining high users of secure messaging within the patient portal using an empirical data-driven strategy (IQR method to identify outliers) that produces contextually based thresholds. We found that a small subset of patients accounted for a relatively large volume of messages (high-volume message senders) and exhibited distinct sociodemographic characteristics and messaging behavior regarding message length, when messages were sent, and messaging use in relation to patient portal use and health care encounters. Measures of messaging intensity were not highly correlated, suggesting that they reflect distinct dimensions of messaging behavior. Recognizing and responding to the different ways patients engage in secure messaging can help health systems enhance patient-clinician electronic communication and enable them to better face the challenges involved in managing patient message volume.