Patient Portal Adoption Barriers in a Safety-Net Clinic: A Mixed-Methods Study With National Comparison

Abstract

Introduction

Patient portals are central to health care communication, yet disparities in adoption persist among underserved populations. Most implementation strategies draw on general population research, but adoption mechanisms may differ substantially in safety-net settings.

Methods

This mixed-methods cross-sectional study examined portal adoption barriers at a student-run free clinic serving predominantly uninsured, Spanish-speaking adults (n = 112 patients, N = 42 practitioners), with comparison to the Health Information National Trends Survey 6 (n = 5232). Exploratory cluster analysis identified patient engagement patterns. Semistructured interviews (n = 11) contextualized quantitative findings. Logistic regression models compared behavioral predictors across settings.

Results

Portal adoption was 22.3% locally vs 66.1% nationally. Cluster analysis identified 4 engagement patterns reflecting distinct barriers, including patients with high health confidence but low information access. Qualitative interviews revealed system usability and trust barriers rather than health literacy deficits. A logistic regression model using behavioral predictors performed well nationally (area under the curve = 0.78) but poorly locally (area under the curve = 0.37), suggesting different adoption mechanisms in safety-net settings. Only 44.3% of practitioners found the portal easy to use.

Conclusions

Equitable digital health implementation requires community-specific barrier assessment rather than reliance on generalized national models. System-level usability and trust barriers, rather than individual behavioral deficits, may drive low portal adoption in underserved populations.

Introduction

Digital health technologies have become central to patient-clinician communication. Patient portals serve as a foundation for health care communication, providing patients with access to health information, secure messaging with care teams, and self-management tools. ^1,2 National policies, including the 21st Century Cures Act mandating patient access to electronic health information, have accelerated portal adoption across health systems. ^3,4

However, the promise of digital health is tempered by the chance of widening the “digital divide.” Despite growing portal adoption, substantial disparities persist, particularly among racially and ethnically minoritized groups, low-income individuals, and those with limited English proficiency. ^5–12 These disparities are not merely about technology access; they are intertwined with social determinants of health, digital literacy, and institutional trust. ^13–18 Identifying the drivers of patient engagement in underserved populations is therefore essential to achieving digital health equity. ¹⁹

Most patient portal implementation strategies draw on general population research, yet the mechanisms driving adoption may differ substantially in safety-net populations where patients face intersecting barriers including poverty, limited English proficiency, immigration-related concerns, and unfamiliarity with digital health systems. ^20–24 To address this gap, the authors employed a mixed-methods approach with 3 aims ¹ : characterize patterns of portal-related engagement and barriers through exploratory cluster analysis, ² contextualize these patterns with qualitative patient and practitioner data, ³ and explore whether established behavioral predictors of portal use in a national sample apply in this safety-net context.

Methods

STUDY DESIGN AND SETTING

This mixed-methods cross-sectional study was conducted at a weekly student-run free clinic (hereafter “the clinic”) serving approximately between 120 and 150 uninsured, predominantly Spanish-speaking adults monthly. The clinic meets national safety-net clinic criteria, with more than 85% of its patients living at or below 200% of the federal poverty level, aligning its patient population with Health Resources and Services Administration–designated health centers. The clinic is staffed by medical student volunteers under faculty supervision, with services including primary care, chronic disease management, and social services referrals. Patients are predominantly undocumented immigrants who lack employer-sponsored insurance and are ineligible for most public coverage programs. Data were collected from October 29, 2022, to May 11, 2024.

ETHICAL APPROVAL AND CONSENT

Survey data were initially collected as part of ongoing clinic quality-improvement activities; the California Northstate University Institutional Review Board (IRB) subsequently determined the study to be exempt (IRB# 2403-02-149) under 45 CFR 46.104(d)(2)(ii), as the survey instrument posed no more than minimal risk. Participants were informed that participation was voluntary and anonymous; given the exempt, anonymous design, written documentation of consent was not required. Participants were also informed that there would be no compensation for participation nor any consequences for declining.

PARTICIPANTS AND RECRUITMENT

The authors analyzed data from the Health Information National Trends Survey 6 (HINTS6), a nationally representative 2022 survey of US adults (N = 6252). ²⁵ The authors excluded participants who had not visited a health care practitioner in the past 12 months (n = 629, 10.1%), those with missing data on key variables (n = 326, 5.2%), and those with question response errors (n = 65, 1.0%). The final analytic sample included 5232 participants, representing 83.7% of the original sample. Of these, 3460 (66.1%) reported accessing an online patient portal. Clinic recruitment flow is shown in Figure 1.

Figure 1:

Patient recruitment and enrollment flow diagram. Consolidated Standards of Reporting Trials–style flow diagram showing the recruitment process at the clinic (October 2022 to May 2024). Of approximately 1400 unique patients served during the study period, 112 participants completed the patient survey on portal use and health information management confidence. Simultaneously, 42 clinic practitioners completed a separate survey on portal implementation. HINTS6 = Health Information National Trends Survey 6.

DATA COLLECTION AND VARIABLES

The authors administered an anonymous, cross-sectional survey (under 5 minutes) via Google Forms, with trained bilingual interviewers assisting patients as needed. The questionnaire measured self-management self-efficacy using 5 items. Three used 5-point Likert scales (1 = strongly disagree, 5 = strongly agree): "I understand my medical condition(s) and symptoms," "I understand my overall health and treatment goals," and "I understand my health information." Two used binary responses: "I can easily access my medical information" and "I can easily share my medical care information with my family/caregivers." Portal use was assessed with a single yes/no item and technology access by smartphone or computer ownership. Spanish versions were forward- and back-translated by 2 bilingual staff members, with adjudication by a third. Interviewers completed a 15-minute training on neutral phrasing, confidentiality, and anonymity. No identifying information was collected. Based on clinic census data over the 18.5-month period, the authors estimated approximately between 1300 and 1500 unique patients, yielding an estimated response rate of 7% to 9% for the 112-patient analytic sample.

Qualitative Methods

Eleven semistructured interviews (9 portal users, 2 nonusers) were conducted using a 10-item guide developed by the coauthors and reviewed by student-volunteer leadership for clarity and cultural appropriateness. Trained undergraduate volunteers administered questions verbatim in English or Spanish in a private clinic setting, transcribing responses onto structured forms.

Transcripts were uploaded to Taguette for thematic analysis using a dual-coder approach. Two researchers trained in qualitative methods independently read the initial 3 transcripts, performed open coding, and then compared codes and resolved discrepancies through consensus to develop a unified codebook with operational definitions. Both researchers independently coded all 11 transcripts. Intercoder agreement was evaluated through systematic consensus-building discussions, with all discrepancies resolved until 100% agreement. Final codes were grouped into higher-level themes, with audit trails maintained throughout. This study followed the Consolidated Criteria for Reporting on Qualitative Research checklist.

SAMPLE SIZE AND POWER

As an exploratory study, no a priori power calculation was performed; the authors enrolled all eligible consenting clinic attendees (n = 112). A posthoc 1-sample Z-test comparing the observed adoption rate (22.3%, 25/112) to the national benchmark (66.1%) yielded greater than 99% power. This applies only to the proportion test; predictive modeling and between-group comparisons should be interpreted as exploratory given the modest sample size (see Supplemental Table 1, which shows behavioral differences between clinic users and nonusers).

STATISTICAL ANALYSIS

Cluster Analysis

All 5 survey items on health understanding and information access were standardized (z scores). Cluster stability was assessed via 1000 bootstrap resamples. The authors applied k-means clustering ^26,27 and selected the optimal cluster number (k = 4) by triangulating 3 metrics: the elbow method, average silhouette width (silhouette coefficient = 0.31), and gap statistic (see Supplemental Figure 1, which shows the elbow method validation). ²⁷ Given the modest sample size, cluster results are interpreted as exploratory patient engagement patterns rather than definitive typologies.

Predictive Modeling

To explore whether standard behavioral predictors apply in this setting, the authors developed a logistic regression model for portal use, selected for its interpretability, appropriateness for binary outcomes, and fit to modest sample sizes. The model was evaluated using 5-fold crossvalidation. Posthoc explainability analyses were not conducted on the local model given its poor discriminative performance.

National Comparison

The authors analyzed the publicly available HINTS6 dataset (n = 5232; see Supplemental Table 3, which shows the demographic breakdown of the HINTS6 sample), aligning key constructs (online medical record access, self-efficacy items) and comparing distributions using Mann-Whitney U tests and Spearman correlations. Dataset factorability was confirmed (Kaiser-Meyer-Olkin = 0.747; Bartlett χ² = 13,740.3; P < .001). ²¹ Classification models were trained using 7 behavioral features with 5-fold crossvalidation using the same framework as the local analysis (see Supplemental Table 4, which shows classification model performance metrics). This comparison is intended to contextualize the local findings rather than serve as formal external validation, given differences in sampling, population, and measurement between the 2 datasets.

Variables were operationalized by analytic method: binary indicators for predictive modeling (clinically meaningful thresholds) and original response scales for clustering (preserving engagement variation).

All analyses used Python version 3.12 (pandas, numpy, scikit-learn, scipy, statsmodels). Statistical significance was 2-sided at P < .05. This study followed Strengthening the Reporting of Observational Studies in Epidemiology guidelines.

Results

LOW ADOPTION AND HIGH ENGAGEMENT

Of 112 patient respondents, only 25 (22.3%) reported active portal use, compared with the 66.1% national benchmark in HINTS6. Despite low adoption, patients reported high confidence in understanding their health conditions and treatment goals. However, confidence in accessing health care information had the lowest agreement among nonusers (74.7%). Mann-Whitney U tests revealed no statistically significant differences between portal users and nonusers across any health information management items (all P > .05), with small effect sizes (Cohen’s d < 0.2). This contrasts with the significant differences observed in HINTS6 (Table 1).

Table 1:

Digital health equity gap: key differences between underserved safety-net population and national survey data

Measure	Clinic (n = 112)	HINTS6 (n = 5232)	Key takeaway
Portal / online record access	22.3%	66.1%	Substantial adoption gap between settings
Confidence in health information	No significant difference between users and nonusers	Statistically significant difference (P < .001). Users are more confident	High confidence locally (M > 4.5) did not predict portal use, suggesting barriers beyond health literacy
Predictive model performance (AUC)	0.37 (No better than chance)	0.78 (strong discrimination)	Behavioral predictors may operate differently in safety-net settings

AUC, area under the curve; HINTS6, Health Information National Trends Survey 6.

PATIENT ENGAGEMENT PATTERNS

Exploratory k-means clustering of patient survey responses suggested 4 engagement patterns, offering a more nuanced view than a simple user/nonuser dichotomy. The clustering solution showed reasonable stability (bootstrap stability = 0.80) and cohesion (average silhouette score = 0.31), although these results should be interpreted cautiously given the sample size. Patterns are detailed in Table 2 (see also Supplemental Table 2, which shows patient engagement pattern characteristics by cluster).

Table 2:

Patient patterns from clustering analysis: characteristics and targeted intervention strategies

Pattern name	Pattern #	Key characteristics	Potential implication
Health care engaged	1	High confidence across all health-info management domains; frequent portal users	Leverage as peer mentors or beta-testers
Digitally disengaged	2	Primarily nonportal users; low ratings on portal features (secure messaging, labs, summaries)	Target for basic digital literacy and adoption
Information access challenged	3	Specifically low confidence/access in retrieving medical information; mixed portal use	Focus on simplifying info-retrieval workflows
System navigators	4	Similar to pattern 1 in most respects, comfortable with systems but not distinguished from health care engaged	Engage as codesigners for portal refinements

• Pattern 1 (health care engaged): high confidence across all domains and frequent portal use

• Pattern 2 (digitally disengaged): predominantly nonusers with low confidence in managing health information; qualitative interviews with patients in this pattern revealed discomfort with technology; a nonuser described the portal as “too hard” to navigate

• Pattern 3 (information access challenged): high confidence in understanding health but low confidence in accessing medical information, suggesting patients who are motivated but encountering usability barriers; qualitative data supported this interpretation; active portal users described difficulty “finding where things are” and called menus “cluttered”

• Pattern 4 (system navigators): moderate engagement and comfort with health care systems, resembling the health care engaged pattern

PRACTITIONER PERSPECTIVES

Surveys of 42 practitioners revealed a disconnect between perceived portal importance and practical usability. Although 88.4% of practitioners found portal content important and relevant for patients, only 44.3% found the portal easy to use from the practitioner side (see Supplemental Table 5, which shows health care practitioner perceptions of patient portal implementation).

EXPLORATORY COMPARISON OF PREDICTIVE MODELS

To explore whether standard behavioral predictors of portal use apply in this setting, the authors built a logistic regression model using patient survey data. The model did not meaningfully distinguish users from nonusers (area under the curve [AUC] = 0.37). Although the small sample likely contributed to this poor performance, the contrast with national data is notable; when the identical modeling approach was applied to the HINTS6 sample, the same behavioral variables were strongly predictive (AUC = 0.78). This divergence is visualized in Figure 2, and different correlation patterns across populations (Figure 3) further illustrate these findings. This pattern is consistent with the interpretation that different mechanisms may drive portal adoption in safety-net vs general populations, although confirmation in larger samples is needed.

Figure 2:

Area under the curve comparison of predictive model performance: local vs national data. Bar chart comparing the predictive performance of identical behavioral models for patient portal adoption. The local safety-net clinic model (n = 112, dark bar) achieved area under the curve = 0.37 (95% confidence interval, 0.28–0.46), performing below random chance (area under the curve = 0.5, dashed line). The national HINTS6 model (n = 5232, light bar) achieved area under the curve = 0.78 (95% confidence interval, 0.76–0.80), suggesting that behavioral factors predictive of portal use nationally may not generalize to safety-net settings. Error bars represent 95% confidence intervals. AUC = area under the curve; HINTS6 = Health Information National Trends Survey 6.

Figure 3:

(A) Health care information management correlation patterns in clinic survey population: heatmap analysis revealing relationship strengths between patient portal adoption factors (n = 112). (B) Health care information management correlation patterns in HINTS6 population (n = 5232). HINTS6 = Health Information National Trends Survey 6.

QUALITATIVE FINDINGS

Interviews with 11 patients revealed 4 themes. Navigation and usability challenges affected both users and nonusers, with patients reporting difficulty "finding where things are" and describing menus as "cluttered." Among users, perceptions of utility were mixed: laboratory results were consistently the most valuable feature, although some users reported "nothing" as particularly useful. Technology comfort appeared linked to adoption decisions, with nonusers consistently reporting greater difficulty ("too hard," "confusing") compared to users who ranged from "easy" to "confusing." Both groups demonstrated awareness of needed improvements, requesting better log in systems and "more organized" interfaces.

Discussion

This mixed-methods study of patient portal adoption at a student-run free clinic serving predominantly uninsured, Spanish-speaking adults found a substantial gap compared with national benchmarks (22.3% vs 66.1%). Qualitative and quantitative findings converged to suggest that system-level usability and trust barriers, rather than individual behavioral deficits, may drive low adoption in this population. A behavioral predictive model that performed well nationally (AUC = 0.78) showed poor discrimination locally (AUC = 0.37), a pattern consistent with different adoption mechanisms operating in safety-net settings, although the small sample warrants cautious interpretation.

The 4 engagement patterns identified through cluster analysis, although exploratory, offer a framework for understanding heterogeneity in portal adoption barriers. The information access challenged pattern is particularly notable; these patients reported high confidence in understanding their health yet low confidence in accessing health information, suggesting that portal design rather than patient capacity may be the limiting factor. The digitally disengaged pattern similarly pointed to the absence of culturally appropriate onboarding support rather than individual deficits. ²⁸

Comparison with national data reinforces this interpretation. In HINTS6, confidence in understanding health information strongly predicted portal use (P < .001). In the authors’ clinic, this same confidence was uniformly high (M > 4.5) across users and nonusers, yet predicted nothing (all P > .05). This suggests that among patients in this setting, health comprehension is not the barrier; rather, system usability and access may be the primary constraints. ²⁹

Qualitative findings converged with the clustering results, with navigation difficulties reported by interviewees directly mapping onto the information access challenged pattern and technology discomfort mapping onto the digitally disengaged pattern.

These findings have implications for both system design and implementation support. The predominance of usability barriers suggests that health care organizations should consider redesigning portals before deploying digital literacy interventions. User experience research ^30,31 identifies cognitive load, navigation complexity, and unclear labeling as strong predictors of technology abandonment. These patterns were mirrored in the qualitative data, with users describing cluttered menus and difficulty locating information. Evidence-based redesign ³² should prioritize simplified navigation with plain-language labels tested in target populations, action-oriented dashboards centered on high-priority tasks, streamlined failure recovery, and iterative usability testing informed by patient engagement patterns like those identified in this study. The information access challenged patients, who were motivated but encountering design barriers, would benefit from such changes.

Practitioner-side usability barriers further compound this challenge. Only 44.3% of practitioners in the study found the portal easy to use, indicating that system redesign must address both patient- and practitioner-facing interfaces.

Beyond system redesign, human-centered support remains essential. Even optimally designed systems require human support for successful adoption in underserved populations. ³³ Systematic reviews indicate that individually focused interventions are among the most effective strategies for increasing portal use in underserved communities. Two complementary models emerge from the patient engagement patterns.

Digital navigators programs may be well-suited for the digitally disengaged pattern. Community health workers or promotoras ³⁴ providing one-on-one, culturally congruent training may help build technological confidence and address the language and trust barriers that limit portal use in this population. For undocumented and underserved Latino and Latina populations, ³⁵ these navigators must do more than teach button-clicking; they must build trust, address privacy concerns, and bridge cultural gaps between patients and institutional systems. The Los Angeles County Department of Health Services' Health Technology Navigator program ²² offers a promising implementation model. Peer mentorship could leverage the health care engaged pattern. Patients who have successfully adopted the portal offer credible, relatable support that institutional staff cannot replicate. Peer support programs are effective across health care domains for building confidence and promoting technology adoption. A peer mentor sharing language, culture, and socioeconomic background with struggling patients can overcome intimidation and mistrust in ways technical training alone cannot.

STUDY LIMITATIONS

Several limitations warrant consideration. The modest quantitative sample (n = 112, 25 portal users) limited statistical power for subgroup comparisons and constrained predictive model stability. The AUC of 0.37 should be interpreted as a general indicator of poor discrimination rather than a precise estimate, and both the cluster analysis and predictive modeling results are exploratory.

The estimated response rate was low (7%–9%), reflecting recruitment challenges in walk-in safety-net settings where patients face competing demands and varying comfort with research participation. This raised selection bias concerns; nonrespondents may have been the most digitally marginalized, potentially causing the findings to underestimate the true extent of portal barriers. These recruitment challenges are themselves informative, reflecting the institutional trust and access barriers that characterize research in safety-net settings.

This study also lacked individual-level demographic data, preventing adjustment for key confounders (age, education, income). This decision was guided by IRB considerations to prioritize participant trust and reduce participation barriers in this underserved community. Future studies with sufficient trust-building infrastructure may be able to collect demographic data while maintaining participation rates.

Conducting this study in a student-run, free clinic also prevented fully disentangling the patient population from the clinic operational model. Future multisite research comparing student-run and professional-run clinics could address this. The cross-sectional design limited causal inferences, and high baseline ratings on many items may have introduced ceiling effects.

FUTURE RESEARCH DIRECTIONS

The authors’ findings suggest several research priorities: evaluating digital navigator programs for the digitally disengaged, testing simplified interfaces for the information access challenged, assessing peer mentorship leveraging the health care engaged, and developing culturally attuned assessment tools for Spanish-speaking populations.

At the policy level, federal digital health equity assessments should extend beyond adoption metrics to incorporate community-specific barrier assessment. At the health system level, implementation strategies could benefit from phased approaches informed by local engagement patterns. At the practice level, trained digital health navigators and partnerships with community-based organizations may be important components of equitable implementation in safety-net settings.

Conclusion

Patient portal implementation in underserved communities requires understanding of community-specific barriers. These findings suggested that uniform digital health approaches may primarily reach already-engaged patients while leaving the most marginalized behind. The divergence between local and national findings (AUC 0.37 vs 0.78) suggested that digital health equity may not be achievable through standardized top-down interventions alone. Health care organizations should invest in understanding community-specific barriers and developing tailored strategies from basic digital literacy support for the digitally disengaged to simplified workflows for the information access challenged. Only through such context-aware approaches can digital health tools become a bridge rather than a barrier for underserved patients.

Supplementary Material

online supplementary file 1: