We identified individuals with ophthalmic disease using the International Classification of Diseases, 10th Revision (ICD-10), codes [8] listed in Table S1 in Multimedia Appendix 1. We extracted a total of 52,388 deidentified, secure messages exchanged through the patient portal of a large tertiary academic hospital, Stanford Health Care, over a period of 10 years (June 2014 to July 2024). We excluded 18,126 (34.6%) messages, including questionnaires, courtesy messages, and physician responses, and 3872 (7.4%) messages routed to nonophthalmology specialties, as shown in Figure S1 in Multimedia Appendix 1. This resulted in 30,390 (58%) patient medical advice requests (PMARs) extracted from the original database of 52,388 secure messages, which were routed to 13 ophthalmology clinics affiliated with Stanford Health Care in Northern California.

To understand patients’ concerns and issues, we developed an NLP topic model using BERT (bidirectional encoder representations from transformers), as illustrated in Figure S1 in Multimedia Appendix 1. First, we preprocessed the raw message data by removing duplicates, special characters, less meaningful messages (eg, less than 50 characters, including “thank you”), and stopwords using CountVectorizer from the scikit-learn library. Second, we applied a widely used pretrained and distilled BERT model (all-miniLM-L6-v2) to convert each message to sentence-level embeddings, which was subsequently simplified by dimension reduction via Uniform Mapping and Approximation and Projection (UMAP). Then, we created a list of topics, applying zero-shot clustering by setting the topic similarity threshold of 0.82, generally considered a good score [9]. Using this threshold, approximately 16,106 (53%) of the 30,390 messages were not assigned to a defined topic cluster, reflecting conservative topic assignment to maximize semantic coherence, which left 14,284 (47%) messages for topic characterization and subsequent analyses. Messages that did not meet the similarity threshold were excluded from downstream topic-specific analyses. Each topic was plotted in the visualization to show the topics by size for quantification and location for relationships between topics.

Each topic was characterized by its 15 most representative keywords. Two investigators (JK and ZZF) manually reviewed the topics and keywords and grouped similar topics into a larger topic. For enhanced interpretability, we applied a large language model (LLM), OpenAI o4-mini (OpenAI), to elaborate on patients’ issues in the HIPAA (Health Insurance Portability and Accountability Act)–compliant analytics environment. As shown in Figure 1, multiple prompt-engineering techniques were used, including role prompting (eg, “you are a professional ophthalmologist in the hospital”), directive commanding (eg, “help interpreting the topic using keywords”), and expertise emulation (eg, “I am also an ophthalmologist, and I will use your response”).

To identify patient groups who would need additional support for care, we computed the rate difference in topics addressed through patient messages by sociodemographic characteristics: age (under 50 years vs 50 years and older), sex (female vs male), race (White vs non-White, including Asian, Black/African American, Native American, Pacific Islander, other, and unknown), ethnicity (Hispanic vs non-Hispanic), and marital status (married vs unmarried, including single, divorced, separated, widowed, life partner, and other). The Rate_A was calculated such that the number of patients in subgroup A who had sent messages on topic X was divided by the total number of patients in subgroup A as computed below.

Because individual patients might have sent messages that were then grouped under multiple topics, these topic-specific rates were calculated independently and were therefore not expected to sum to 100% within a demographic subgroup. The statistical significance of the rate difference was determined by a 2-proportional z-test when P<.05. To visualize the relationships of the age and race categories with the patient issues, we plotted a Sankey diagram using detailed age (18 to <35 years, 35 to <50 years, 50 to <65 years, 65 to <75 years, and 75 years and older) and race (Asian, Black, Native American/Pacific Islander, other, and White) categories. We used Python 3.10 in Colab Enterprise (Google).

This study was approved by the Stanford University Institutional Review Board (IRB number 71157). The requirement for informed consent was waived because the study involved a retrospective analysis of deidentified patient data, posed minimal risk to participants, and could not practicably be conducted with individual consent. No direct patient contact occurred. All data were analyzed in a secure, HIPAA-compliant environment, and no personally identifiable information was included in the analysis or presented in this manuscript. Our research follows the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines.

Our findings should be interpreted in the context of prior studies examining ophthalmology-related questions on public online platforms. Previous analyses of online Reddit and EyeTriage forum posts demonstrated that patients frequently seek information regarding ocular symptoms, treatment options, and urgent eye concerns outside traditional clinical settings [10,11]. Unlike these studies, however, our study evaluated secure patient portal messages exchanged directly between patients and their clinical care teams within the EHR system. As a result, our dataset captures real-world clinical and administrative concerns linked to ongoing care, allowing examination of demographic differences in messaging patterns across a large ophthalmology population. To our knowledge, this study represents one of the largest analyses of secure patient-provider ophthalmology messaging within an integrated EHR system and provides insights into patient communication patterns, topic distributions, and demographic differences in portal-based ophthalmic care. We found that scheduling requests and vision-related concerns were the most frequent message topics, alongside glaucoma and surgical questions, also highlighting subtle but important disparities in who uses these digital services. These insights provide a roadmap for improving digital ophthalmic care delivery. This analysis of over 30,000 ophthalmology-related messages received over a 10-year span on the patient portal reveals clear demographic patterns, distinct thematic trends in patient concerns, and actionable opportunities for AI-assisted triage. By linking large-scale NLP topic modeling with sociodemographic analysis, our findings highlight both the promise and the challenges of integrating digital messaging into ophthalmic care. Knowing the most common message topics (and which topics often signal urgent issues) allows the development of AI triage algorithms that flag critical ocular symptoms for prompt attention. Our data could inform training of an NLP system to distinguish routine questions from red-flag symptoms, such as sudden vision loss, that require a same-day response.

More than half of the message senders were female (2673/4817, 55.5%), and more than half were aged 50 years or older (2740/4817, 56.9%), consistent with the demographic profile of many ophthalmic conditions [12,13]. The predominance of women aligns with prior digital health literature suggesting higher female engagement with patient portals, possibly due to greater health-seeking behavior, higher prevalence of certain eye conditions, or the convenience of remote messaging for those with mobility or caregiving responsibilities [14]. Racial and ethnic distributions reveal potential access dynamics: although White patients comprised the largest racial group, non-White patients accounted for a slightly greater proportion of total message senders than seen in traditional in person visit patterns [12]. Notably, non-White patients more frequently messaged about glaucoma, pharmacy refills, insurance, and disability, while White patients disproportionately messaged about surgical issues, patterns that likely reflect known differences in disease burden, insurance complexity, and procedure uptake [15].

The finding that non-White patients raised more administrative and chronic disease management concerns underscores existing disparities in health care navigation and access [15]. Barriers such as insurance authorization complexity and pharmacy access may disproportionately affect these groups, suggesting that portal-based AI tools could be used to streamline such processes. Conversely, the higher rate of surgery-related messages from White patients may reflect greater access to elective or nonurgent procedures, emphasizing the need to monitor whether digital tools inadvertently reinforce existing inequities [13].

While older patients were not underrepresented in portal use, they contributed proportionally more clinical queries (particularly on glaucoma, surgery, and tumor imaging) than younger patients, highlighting age-related differences in how the portal is used. This likely reflects both the higher prevalence of chronic, vision-threatening conditions such as glaucoma, cataract, and ocular tumors in older adults and their greater need for perioperative communication given the frequency of surgery in this age group. In contrast, younger patients may rely on in person visits or alternate communication methods for routine concerns, resulting in fewer portal-based clinical queries. These findings align with emerging evidence that older adults increasingly engage with digital health technologies. Recent national survey data demonstrated substantial use of personal EHRs among older adults, particularly among those with greater health care needs [16,17]. Our findings extend this literature by suggesting that older adults not only access patient portals but also use them as a clinically meaningful communication channel for managing chronic ophthalmic conditions and treatment-related concerns.

While nearly half of all messages involved administrative matters such as scheduling and refills, clinical issues such as vision disturbances, glaucoma, and surgery-related concerns were highly represented. The strong presence of chronic disease topics such as glaucoma (particularly among older and non-White patients) reinforces the potential of patient messaging as an early touchpoint for ongoing disease monitoring, adherence reinforcement, and symptom triage [18,19].

Given that many high-volume, low-complexity message types (eg, scheduling, refills, insurance verification) could be automated, AI-driven triage systems could significantly reduce clinician burden [2,4]. Our NLP approach successfully identified topical categories, suggesting feasibility for real-time routing of urgent messages (eg, acute vision changes) to clinicians while directing routine queries to automated workflows [5,7]. Previous studies show that LLMs can approach ophthalmology trainee accuracy in triage and diagnosis, but accuracy, bias mitigation, and safety require further validation in live settings [20]. The variation in topics by demographic group also underscores the importance of training and testing AI systems on diverse datasets to avoid perpetuating disparities [21].

An AI-enabled portal triage system could flag urgent ophthalmic complaints for expedited review, automate routine administrative responses, and support multilingual or literacy-sensitive messaging for patients at risk of being underserved [22-24]. Integration with EHR alerts, automated follow-up prompts, and links to teleophthalmology services could further extend the reach of such systems, especially in areas with limited specialist access [19,25].

This study was conducted within a large academic health care system in Northern California and may therefore reflect the demographic composition, referral patterns, and digital health infrastructure of this region. Patients using secure messaging may differ from those receiving care in community-based, rural, or underserved health care settings. Furthermore, disparities in internet access, digital literacy, language proficiency, socioeconomic status, and patient portal enrollment may have influenced messaging behavior and contributed to observed demographic differences. These factors should be considered when generalizing the findings to broader populations and when developing equitable AI-assisted triage systems. Despite these limitations, the large sample and clear patterns support the main conclusions of this study.

Additionally, nearly half of the messages were not assigned to predefined topic clusters due to the conservative similarity threshold used in zero-shot clustering. While this approach improves topic specificity, it may limit completeness. Future work should explore lower similarity thresholds or supervised classification methods to improve coverage and classification accuracy. Similar analyses could benefit any field where patient portals are used, and our approach demonstrates how mining patient messages can identify unmet needs and system inefficiencies. The predominance of administrative and nonacute clinical messages, such as medication refills, dry eye symptoms, and postsurgical follow-ups, suggests that many patient concerns may be suitable for remote management or asynchronous triage, although message urgency was not formally analyzed in this study. While our analysis characterizes the content and demographic patterns of patient messages, we were unable to determine how each message was triaged in practice, whether it prompted an appointment, a same-day urgent visit, or an asynchronous reply. Because message outcomes were not linked to clinical scheduling or chart data, the urgency of individual threads could not be independently verified. This limitation should be considered when interpreting the potential of NLP tools for real-world triage, as their clinical utility ultimately depends on validated outcome mapping.

The study period spanned a decade from 2014 to 2024, encompassing substantial changes in health care delivery, patient portal adoption, and digital health technology, including the COVID-19 pandemic, which produced a marked and durable increase in secure messaging volume across health systems [1,26]. Message volume, topic distribution, and portal engagement may therefore have varied across prepandemic, pandemic, and postpandemic periods. Our topic model was derived from the pooled decade-long corpus to maximize semantic coherence and cluster stability; period-stratified rederivation of topics, which would be required to compare content across eras without cluster contamination, was outside the scope of this characterization and represents an important direction for future work. Examining how message content and the demographic composition of senders shifted across pandemic periods may further refine the design and ongoing calibration of AI-assisted triage tools, which are sensitive to temporal drift in input data. Building on these findings, future work should examine whether portal use correlates with downstream outcomes such as appointment adherence, surgical uptake, or reduced emergency department visits [26,27]. Prospective evaluation of AI triage performance in real-world ophthalmology workflows will be critical. Furthermore, targeted interventions (such as insurance navigation support for non-White patients and early glaucoma education prompts for high-risk groups) could be tested to assess their impact on health equity and clinical outcomes.

Secure patient messaging, enhanced by NLP, reveals meaningful patterns in ophthalmic concerns and disparities in care-seeking behaviors. Non-White patients more often raised issues related to insurance, pharmacy refills, and chronic disease management, while White patients focused more on surgical topics, underscoring opportunities for targeted interventions. The prevalence of high-volume, low-complexity queries supports integrating AI-assisted triage into patient portals to prioritize urgent complaints, automate routine responses, and improve multilingual, literacy-sensitive communication. Such systems have the potential to enhance efficiency, equity, and timely access to specialty care but must be validated for accuracy, safety, and bias mitigation before widespread implementation. These findings support the development of AI-driven triage systems using real-world patient-generated data in EHR portals. By aligning AI-driven triage with equity-focused design, patient portals can transform from passive communication tools into proactive, data-driven components of ophthalmic care delivery.