This study was found exempt by our institution’s institutional review board. To gather large-scale online posts by women with UTI, we contracted with Treato, a web-based data-mining service company that utilizes extraction templates and a proprietary search algorithm designed to capture patient content. After consultation with the PLUS Consortium, our team chose a combination of keywords related to disease nomenclature, symptoms, treatment options, and exclusion terms to identify posts using their search algorithm (Multimedia Appendix 1). After extracting posts from online forums, we performed digital ethnography using qualitative thematic analysis and a latent Dirichlet allocation (LDA) topic modeling quantitative process that also facilitated qualitative results [11]. Combining both methods allowed us to ensure thematic saturation by analyzing the entire data set of identified posts.

After identifying posts that met our search criteria, the entire data set was randomized to ensure that we reviewed posts from various websites before reaching thematic saturation. We performed an inductive, iterative, open-coding, qualitative analysis until we could no longer identify unique themes. Two research team members were assigned to examine the extracted posts several times. The data were then organized into different units or codes, which provides sufficient detail for the reader even without the context. Therefore, the codes were supported by text fragments. This was an iterative process; hence, as unique inductive codes emerged, they were regrouped into more specific categories, and some were combined while others were placed in a superordinate category. Our goal was to avoid redundancy among the categories, so we created broad themes encompassing the categories.

To supplement the manual inductive coding process, we applied a second, more novel technique, LDA, that allowed for the review of the entire data set. LDA is an unsupervised probabilistic topic model process that relies on the contextual co-occurrence of words to identify patterns of words that, when found together, have a semantic meaning [11,12]. For example, the word “bank” can have different meanings when paired with “money” versus “water.” This model generates outputs as topics that can be understood with the concept of a “word cloud,” comprised of words that are ranked higher in a corresponding topic, if these co-occurred frequently in the social media posts. These topics were interpreted for thematic analysis. Each topic has an assigned prevalence value, which represents the quantity of words in the collection assigned to the topic divided by the total number of words. The word cloud topics were sorted based on their respective prevalence (quantitative signature) for review by the research team to identify a thematic interpretation. Consistency of theme allocation for a specific topic was confirmed by reviewing posts that contributed to the development of the word cloud. Table 1 demonstrates examples of the word cloud topics and their assigned themes. Combining both methods allowed for a comprehensive review of the results using a semiautomated approach and a manual inductive coding process to capture a broad understanding of the experience of patients with UTI.

We identified a total of 105 themes using LDA, which were grouped into 9 categories to avoid redundancy and provide an overview of the topics represented online (Table 2). Additionally, there was significant overlap with our hand-coding approach, so the data were synthesized into 7 themes with subthemes (Table 2) to represent results from both methods. Our hand-coding approach facilitated more descriptive interpretations.

Qualitative hand-coding analysis yielded 7 themes with subthemes related to the knowledge and experience of women with UTI symptoms (Table 3).

Our ethnographic study of social media posts on UTIs revealed information on illness experience, lay knowledge, and concerns among women. Unlike prior qualitative studies, we presented patient perspectives that are likely more diverse and candid than data gathered from specialty clinics [5-7]. We found a strong online community support network created via forums to exchange information among peers, which may have partially resulted from frustrations and challenges with medical care. We found that UTIs cause a significant burden on daily activities and, as a result, women engage in supportive conversations about physician interactions, antibiotics, alternative therapies, risk factors, and prevention strategies to bridge knowledge gaps and obtain reassurance from peers.

We captured broad and diverse patient experiences using two methods to conduct digital ethnography. However, our inductive hand coding provided more granular details as expected from directly analyzing quotes, which helped us comprehend online discussions and women’s perspectives for specific themes. For example, culture-directed antibiotic treatment was a unique patient concern identified with our hand coding of posts that was not found using LDA. Although the LDA word clouds consistently represented quality-of-life concerns, hand coding provided examples of fears women faced. Half of the LDA themes related to community support and identifying risk factors and prevention strategies, which was consistent with our hand-coding results.

Ghouri et al [13] previously conducted 15 telephone interviews with women with prior documented UTI and described feelings of hopelessness and lack of support in this group. Our study highlights the support experienced by patients with UTI who exchanged information online. Our findings support prior results that social media is an integral part of processing medical information and can facilitate patient engagement for the exchange of condition-specific knowledge [14,15]. Additionally, it has been widely documented that online forum discussions allow for better intake and information processing [16-18].

Our study was broad, capturing different populations of women. This better allowed the analysis to be guided by the direct, anonymous discussions of patients, making it more likely to be generalizable to the UTI population at large. Prior work only surveyed patients recruited from clinical settings, those with rUTI, and pregnant women [5-7]. Our work, on the other hand, captured the perspectives of several populations in a single analysis, including pregnant women, geriatric women, and women with rUTI across a collection of websites. Geriatric patients, pregnant women, older women, and women with rUTIs were frequently identified in posts with unique concerns, suggesting the need for more targeted outreach to these special populations.

Unlike prior online studies, our study design has the advantage of analyzing multiple websites [5,6]. Flower et al [5] conducted an analysis of 1 online self-help forum for patients with rUTI to understand how women manage their rUTIs. Our findings of alternative therapies, antibiotic concerns, and patient-physician interactions were similar. However, due to the broader sample size, we found other alternative therapies not previously described, additional antibiotic concerns (eg, treatment duration, human microbiome, bacterial resistance, and culture-guided treatment), and more complex health care barriers (eg, specialty care access, insurance coverage, and presumed level of care required for treatment). To our knowledge, the theme of risk factors and prevention has not been previously described in prior UTI online forum literature. Additionally, our findings that pregnant women were concerned about the progression of cystitis to pyelonephritis, as well as the effect of antibiotic use (or lack of treatment) on fetal development, were consistent with those previously found in a study analyzing online content to understand UTIs and antibiotic use in the pregnant population [6].

The semistructured interview style of many qualitative studies may limit and potentially narrow the scope of what patients share in clinical settings. One qualitative one-on-one interview study of 21 women recruited from a larger primary care trial found that patients wanted clinicians to address quality-of-life impact and that they were receptive to the strategy of antibiotic delay, which allows for 48 hours to reassess if infection symptoms subside before starting antibiotics [7]. Women who enrolled in this randomized study were, by default, receptive to different management strategies. Although we found similar quality-of-life concerns, we also identified self-blame, mismanagement in the primary care setting, delayed referrals, inconsistent counseling about treatment guidelines, unmet expectations, and the practice of culture-directed antibiotic use, which were not previously characterized.

Online discussion points were in agreement with the 2019 American Urological Association’s guidelines for uncomplicated rUTI, which recommend first-line antibiotic agents and promote culture-directed antibiotic treatments rather than empiric treatment [19]. This suggests that many patients with UTIs were well educated on the topic. However, there were inconsistencies discussed online for the role of cystoscopy and upper tract imaging, despite the recommendations provided in the updated guidelines to avoid those diagnostic studies in uncomplicated cases [19]. This may have been due to the fact that our data were collected before the publication of the updated guidelines. Concerns about antibiotic collateral damage mentioned in the literature and discussed in the rUTI guidelines were also supported by our analysis [19].

Despite the innovation and patient inclusivity of our study, there are important limitations that can inform future work. We did not have access to demographic information, and the website content could have been restricted by the sample of websites accessed by Treato and our search strategy. Although we focused our analysis on women, it is possible that some men participated on the forums and were included in the analyses. The anonymous data may also contribute to patient misclassification since we cannot confirm a diagnosis, but our best attempt was made using contextual factors. Additionally, our analysis and conclusion relied on the degree to which individuals post online. We could not characterize specific posts’ engagement level, such as individual read and reply counts. Our study, by default, excluded those women who do not exchange medical information on the internet.

Digital ethnography combining qualitative analysis and LDA allowed us to analyze free-range patient perspectives, which are currently not found in the UTI literature. First, unlike focus group studies, anonymity is a clear driver of candid, honest conversations, facilitating online users to provide support and address the most important concerns. Second, there was a pervasive element of fear: fear of not treating UTIs, as well as fear of the sequelae associated with antibiotic treatments. Finally, the use of online forums empowered women to self-manage their condition and take their care into their own hands. Our findings also demonstrate the reliability of using online social media data to learn about patient behavior and decision-making, which is important to guide how we engage with patients and disseminate society-sponsored guidelines. Patient information, outreach, and treatment guidelines by medical societies must be congruent with patients’ concerns. Physicians can use this data to discuss misconceptions and improve patient-centered care.