This was a secondary analysis of data from a randomized controlled trial that tested a 30-day automated SMS text messaging intervention among primary care patients after hospital discharge (ClinicalTrials.gov NCT05245773). We analyzed SMS text messages, patterns of engagement, and clinical data from patients enrolled in the intervention and who responded to any messages. Patients eligible for the original study were adults (aged 18 y or older) who received care in 30 primary care practices within the University of Pennsylvania Health System; discharged to home from an acute care hospital; and identified as medium to high risk at the time of discharge (using an Epic Systems Corporation developed and validated point score based on clinical information presented in prior literature and generally available in the electronic health record [EHR]; see Multimedia Appendix 1 for further details) [13,14]. The SMS text messaging program was built on Way to Health, a platform created with National Institutes of Health funding to provide automated technology infrastructure in support of clinical care and care delivery innovation research [15]. Figure 1 shows the workflow from primary data collection through feature engineering, clustering, and characterization of phenotypes.

Patients in the original study were randomized 1:1 to either usual transitional care (a single phone call from the practice within 2 business days of discharge) or usual care plus receipt of a 30-day automated SMS text messaging program. The study operated under a waiver of informed consent, though patients were free to opt out via text at any time. A total of 2352 patients were included in the intervention arm.

SMS text messages were sent out on a tapering schedule after discharge. A typical message asked, “Is there anything we can help you with today?” If they answered no, there was no further action. If they answered yes, a follow-up message asked them to further categorize their need (eg, “I need help with my medicines”). Patients were also told to reach out anytime outside of a scheduled check-in context by texting “Call.” Patients who reported a need through either of the methods above (an “escalation,” which was routed to an EHR in-basket) would receive a follow-up phone call from the practice within 1 business day. See the original study for a full schedule and script of all messages [11].

Our automated SMS text messaging program applied prespecified rules to incoming messages from patients. For instance, when patients were prompted with “Is there anything we can help you with today?” the software accepted only a narrow set of responses (Y, Yes, N, or No). Despite this, patients could respond as desired without a limit on character count. Messages containing text lacking a prespecified response triggered an outbound SMS text message to patients stating, “I don’t understand that response. Valid choices are: [prespecified choices].”

The original study was reviewed and approved by the University of Pennsylvania Institutional Review Board (IRB) (Protocol: 849348). The requirement for informed consent was waived for the original study with approval of the IRB; participants were free to opt out of the text messaging program at any time. No compensation was provided to participants.

As a secondary analysis, this study was reviewed and deemed exempt by the University of Pennsylvania Institutional Review Board (Protocol: 855553). The original IRB approval covered secondary analyses, and the waiver of consent similarly applied. All data were deidentified.

A total of 1731 patients engaged with the messaging intervention and were included in the analysis, among which 1060 (61.2%) patients were female; the mean age was 64.8 (SD 16.1) years; 782 (45.2%) and 828 (47.8%) patients identified as Black and White, respectively; and 970 (56%) and 317 (18.3%) patients were insured by Medicare and Medicaid, respectively (Table 2). The mean hospital length of stay was 4.2 (SD 4.6) days, the Charlson Comorbidity Index was 4.7 (SD 2.9), and the LACE score was 66.7 (SD 13.8).

Applying the k-means clustering method to the full set of features, we observed 4 distinct clusters of patient interaction phenotypes. Measures for each feature, stratified by cluster, can be found in Table 3. The enthusiast (n=1029) was characterized by both high engagement (responding to 91.6% of messages, for instance) and high conformity (a 4.6% error rate). The minimalist (n=515) was characterized by low engagement (response rate of 35.2%) but high conformity (4.4% error rate). The nonadapters (n=170) were characterized by both low engagement (39.1% response rate) and conformity (40.8% error rate). Finally, the high-need responder (n=17) was characterized by high engagement (82.4% response rate), moderate conformity (18.5% error rate), and notably, a high rate of inbound messages (3.1 per d) and requests for help (64.7% requesting a call outside of a check-in window).

In Table 2, we observed differences in most demographic characteristics across interaction phenotypes. Nonadapters were more likely to be older (mean age 68.9, SD 14.9) and male (n=79, 46.5%) than other groups. Enthusiasts were more likely to be White (n=570, 55.4%) and have commercial insurance (n=240, 23.3%) than other groups. Minimalists were more likely to be younger (mean age 62.7, SD 17.3) and Black (n=304, 59%) than other groups. There were no major differences in clinical characteristics across phenotypes.

We tested the association between the four interaction phenotypes and the clinical outcome of hospital revisits. In Multimedia Appendix 2, the high-needs responders had significantly higher rates of revisits at 7, 30, and 60 days (35.3%, 52.9%, and 58.8%, respectively). The enthusiasts were the lowest overall at 30 and 60 days (15.8% and 26%, respectively), while the minimalists and nonadapters were similar.

We identified 4 distinct patient interaction phenotypes—enthusiasts, minimalists, nonadapters, and high needs responders—with a postdischarge SMS text messaging program and observed differences in both patient characteristics and outcomes across these groups. Notably, there was an association between high-needs responders and enthusiastic interaction phenotypes and the outcome of hospital revisits.

As the use of mobile health (mHealth) has grown, so has interest in how patients engage with it. Indeed, most studies of mHealth interventions include some measure of user engagement, ranging from app logins to message response rates [16]; however, very few capture the heterogeneity of patient interaction styles. The notion of digital phenotyping, more generally (characterizing interaction with a range of digital health tools), is a nascent but growing field [17]. A similar study of a remote blood pressure management program also characterized engagement styles with automated SMS text messaging specifically, though, notably, this was a different context in both its use case (chronic disease management) and its population (patients who actively consented, vs this study which used a waiver of informed consent and may therefore be more generalizable) [18].

While it is not surprising that patients vary in both the degree to which they engage and their ability to grasp the rules of the system, most programs (including the one studied here) are designed as a one-size-fits-all approach [2]. The notion of just-in-time adaptive interventions has emerged in recent years—with mHealth being a prime use case—which aims to find the right intervention, for the right patient, at the right time [19-21]. This design challenge is related to the one explored here, and in some respects, it considers similar data about the patient. It rests on a slightly different premise about the patient, however, which is that in the course of an intervention, patients may require different strategies to keep them engaged, and we can learn over time how to adjust the timing and intensity of our approach to meet their needs at that particular moment [22].

Behavioral phenotyping asks more fixed questions about patients—for example, is this someone who engages with mHealth technology at all, and if so, how? [23] This is especially important in a population health context, such as ours, where patients are not being recruited for the program but rather automatically enrolled. While we found a high rate of overall engagement in our initial study (79.5% of participants responded to at least 1 message) [11], our analysis here found that this simple measure masks significant heterogeneity in sustained engagement and understanding of program rules. Indeed, as posited by the Technology Acceptance Model [5], perceptions of both usefulness and ease of use may explain why some patients engage intensively while others disengage, underscoring the need for tailored approaches that align more closely with individual patient preferences.

Capturing this information early in the course of an intervention can inform distinct approaches for these groups of patients. For instance, enthusiasts may not need additional intervention. Nonadapters may need extra attention, which could entail additional coaching to help guide their use of the program or a basic needs assessment that may suggest an entirely different medium better suited to their communication preferences. On the other hand, minimalists—who seem to understand the use of the medium (based on low error rates) but simply do not engage—may be best suited for just-in-time adaptive intervention strategies. Although high-needs responders constitute only ~1% of participants, they account for a disproportionate number of 30-day revisits. Flagging patients who generate multiple escalations or unusually frequent messages early in the program would let care managers focus on proactive calls, expedited follow-up visits, and pharmacist reviews on this small but high-risk subgroup.

This can have important implications for digital disparities, as well. In our original study, older adults appeared to engage at comparable rates to the general population, based on a simple measure of engagement; we found here, however, that nonadapters tended to be older, suggesting a need for additional support. We also found that enthusiasts were more likely to be White (n=570, 55.4%), while minimalists were more likely to be Black (n=304, 59%), suggesting the need for alternative strategies to narrow the engagement gap.

Future iterations of postdischarge outreach could layer these phenotypes onto a broader, multimodal ecosystem that matches patient needs and preferences. For instance, a smartphone app that centralizes text, video check-ins, medication reminders, and symptom tracking, or the use of wearable devices that can passively monitor indicators of early deterioration and trigger proactive outreach [24,25]. For patients without smartphones, interactive voice-response calls can be used [26]. An artificial intelligence–driven messaging layer could adjust cadence, channel (text vs voice), and language complexity in real time based on each patient’s engagement phenotype and predicted risk.

Finally, a novel finding of this work is the potential predictive value of users’ interaction styles. While mHealth programs typically rely on information communicated directly through the program (eg, the content of patients’ SMS text messages), our findings suggest that additional information may be gleaned from how patients interact with the program. We found, for instance, that a small cadre of users (high-needs responders) who had very intense communication and needs (several messages per day, high rate of requesting calls) returned to the hospital at a very high rate (52.9% with a 30-day revisit). Conversely, enthusiasts—who had a high response and low error rate—had the lowest 30-day revisit rate. Taken from this perspective, interaction phenotypes can be considered as another input into the overall risk assessment of patients and may help inform who needs additional transitional care support after discharge.

The limitations of this study include that it is a retrospective analysis of a study from a single academic health center, which may limit generalizability to other settings, although it contained several practices and had a diverse patient population. Given the time-limited nature of the original intervention (30 d), there was a limited number of outbound messages, although enough to capture a dispersion of engagement patterns (and patients could text freely outside of the scheduled check-ins). The associations between phenotypes and outcomes were purely descriptive and did not adjust for other potential confounders. Because the high-needs responder group contained only 17 patients (~1% of the cohort), its estimates are inherently imprecise. Finally, while we applied a k-means clustering approach, traditional methods for selecting the optimal number of clusters (elbow method for sum of squared error plots, silhouette score plots) did not converge well on a single answer (3 or 4 both demonstrated arguable change in slope); however, 4 clusters was selected as it identified a distinct subgroup—high needs responders—that was noticeably different from the others. New features may suggest a different k.

This study also had several strengths. It is one of the only efforts to characterize patients’ interaction styles with automated texting. We had access to data from a large trial and were able to analyze a large number of participants. Finally, as noted, the study did not require informed consent, and patients were enrolled automatically if they met the inclusion criteria, making for a more generalizable patient sample.

In this secondary analysis of a postdischarge automated SMS text messaging program, we observed variability in patient engagement. While the initial program may be a scalable template for patient follow-up, our findings suggest two key considerations for future iterations of this approach: (1) some patients may benefit from alternative communication strategies or additional guidance, and (2) patient engagement patterns with the program—in addition to the information they share directly—could offer predictive signals into adverse outcomes. Future research should explore integrating these insights into transitional care messaging programs and digital health programs, more generally, to improve patient engagement and outcomes.