Purposive sampling was used to recruit patients meeting the inclusion criteria: (1) adults with a histologically or cytologically confirmed malignant solid tumor, (2) an education level of at least junior high school, (3) Eastern Cooperative Oncology Group performance status of 0‐1, and (4) estimated life expectancy of 12 weeks or more. These criteria were established to ensure that participants had sufficient physical stamina to complete 20-40–minute interviews and the cognitive capacity to articulate complex factors in clinical trial decision-making. It is acknowledged that these criteria inherently exclude the most frail patients and those with limited educational backgrounds, and that their perspectives may therefore be underrepresented in this study. All participants were required to participate in the study voluntarily, demonstrate adequate comprehension of the study procedures, and provide written informed consent. Patients were excluded if they had a history of psychiatric disorders, cognitive impairment, communication barriers, or active central nervous system metastases (eg, cerebral edema and progressive lesions). After fully explaining the study’s purpose and procedures to all participants, informed consent was obtained. A total of 11 participants were recruited for the face-to-face interviews.

Inclusion criteria are (1) comments explicitly discussing cancer drug clinical trials, (2) comments from patients with cancer or their caregivers, and (3) sufficient content length (≥10 Chinese characters) for meaningful analysis.

Exclusion criteria are (1) duplicate comments, (2) incomplete or incomprehensible content, and (3) advertisements or promotional information.

Face-to-face interviews were conducted in January 2025. All interviews were conducted by a trained qualitative researcher with experience in clinical trials and oncology care. Each patient was interviewed once; no repeat interviews were conducted. Interviews were scheduled at times convenient for the interviewees and conducted one-on-one sessions in quiet conference rooms, lasting 20 to 40 minutes, to allow participants to independently determine the extent of their disclosures. All interviews were audiorecorded with participants’ informed consent.

For this study, online data were collected via web crawlers, and comments published between September 2010 and May 2025 were retrieved. The textual data originated from 2 distinct Chinese online platforms: Zhihu and Yu-Ai-Gong-Wu (Dancing with Cancer). The selection of these 2 platforms was deliberate, aiming to encompass a wide range of viewpoints by combining extensive public discussions with personal accounts that closely address patients’ needs.

Zhihu is a prominent question-and-answer–based social platform in China, known for its high-quality, in-depth content from a diverse user base of professionals, experts, and engaged laypersons. The total number of views on cancer-related topics on Zhihu has exceeded 3.81 billion. Yuaigongwu, which translates to “Dancing with Cancer,” is a dedicated online support community for patients with cancer and their families, established in 2010. The platform facilitates peer-to-peer support and the sharing of firsthand treatment experiences and emotional journeys. The forum has more than 1 million posts and a membership exceeding 180,000.

Data collection was performed using a custom Python web crawler. The crawler was programmed to scrape publicly accessible text data (posts and comments with timestamps) from the selected platforms by targeting specific keywords, including “drug clinical trials” and “oncology drug testing.” All data were anonymized to protect user privacy, and the collection process complied with relevant laws and the platforms’ terms of service. A rigorous multistage data preprocessing protocol was subsequently implemented to ensure the dataset’s quality and relevance.

A total of 3148 comments related to cancer clinical trials were identified through web crawling. After removing 39 duplicate comments, 3109 unique comments were retained for screening. In the initial screening stage, comments were excluded if they (1) contained advertisements or promotional information (n=75), (2) were unrelated to cancer clinical trials (n=2475), or (3) contained fewer than 10 meaningful characters after removing emojis and hyperlinks (n=208). A total of 2758 comments were excluded, and 351 comments were retained for further eligibility assessment.

In the second-stage eligibility assessment, 2 researchers independently reviewed the remaining 351 comments. Comments were excluded if they were unrelated to the subject of this study (n=118) or were incomplete or unintelligible (n=14). Discrepancies between the 2 reviewers were resolved through discussion and, when necessary, adjudicated by a third researcher. After excluding 132 comments, a final dataset of 219 valid comments was retained for analysis.

Text data were normalized by removing nonessential characters, standardizing punctuation, and tokenizing emoticons that conveyed emotion. All inclusion and exclusion decisions were recorded in a structured dataset that included the post ID, screening results, reasons for exclusion, and reviewer notes. A random 5% of the data was rechecked upon completion to ensure accuracy. This 2-step, transparent screening process ensured the reliability, validity, and reproducibility of the dataset used for qualitative analysis.

Content analysis was used to analyze the interview transcripts and online comments separately, thereby distilling key themes. Two researchers independently analyzed the materials by repeatedly reading the textual data, identifying meaningful statements, and assigning codes. Following the summarization, preliminary themes were formulated. In instances of discrepancies during the analysis process, these were resolved through research group discussions until consensus was reached, ultimately determining the final research results.

All interview records were anonymized before analysis to protect confidentiality, especially during the LLM-assisted analysis stage, where special attention was paid to data privacy.

The study required anonymized transcripts to be prepared in plain text format for processing by both Gemini Pro 2.5 (version released by Google AI Studio in March 2025) and DeepSeek R1 (version released in January 2025, accessed via [19]). An initial prompt was developed based on the research objectives and tested on a calibration subset of 20 online comments. The initial prompt used for both LLMs was:

This calibration phase identified two issues requiring refinement: (1) the initial prompt allowed paraphrased summaries, which hindered quote verification, and (2) output formatting was inconsistent across responses. The prompt was refined to require verbatim quotes and to specify a structured output format. Once finalized, the prompt was fixed and applied consistently to all subsequent analyses without further modification.

Unlike human coders, LLMs do not apply predefined codebooks line by line or engage in iterative coding. Consequently, traditional inter-rater reliability metrics, such as Cohen κ, do not apply to their outputs. Instead, transparency in prompt construction and in human review of AI-generated themes was used to ensure analytic consistency and validity [18].

The initial prompt was refined as follows:

These refinements improved the consistency and verifiability of LLM outputs.

Interview transcripts (N=11) were compiled into a single anonymized document with participant identifiers (P01-P11). Online comments (n=219) were organized into a separate document with sequential identifiers (ZH01-ZH53 and YAGW001-YAGW166), with each comment clearly delimited. Documents were uploaded to each LLM platform separately as contextual input, and the models were prompted with structured instructions to identify overarching themes through in-context learning.

LLM outputs were independently reviewed by two researchers to verify (1) accuracy of quote attribution (ie, that quoted text actually appeared in the source data), (2) coherence and relevance of thematic categorizations, and (3) absence of fabricated or hallucinated content. Discrepancies were resolved through discussion with a third researcher. Quotes that could not be verified in the source data were excluded from the final analysis.

To ensure the cross-method reliability of the findings, a systematic comparative analysis was conducted. Themes derived from all analytical methods were evaluated through parallel examination, addressing both terminological variations and interpretive differences. A thematic map was created to link synonymous or related labels, producing a single, integrated thematic framework (Figure 2).

Conceptual alignment was observed when distinct terminologies referenced identical constructs. For instance, the investigator-led subtheme “concerns regarding efficacy,” Gemini Pro 2.5’s “fear stemming from uncertainty,” and DeepSeek R1’s “uncertainty of curative effect” were found to be semantically equivalent and were harmonized into the unified theme of “uncertain therapeutic efficacy.” Conversely, interpretive divergences requiring adjudication occurred when methods captured different intensities or facets of a topic. Rather than treating these as discrepancies, they were leveraged as sources of deeper interpretive insight. A notable discrepancy arose regarding economic motivations: while human analysis identified general “alleviation of economic pressure,” DeepSeek R1 focused on procedural “cost waivers,” and Gemini Pro 2.5 highlighted a more desperate sentiment that “free treatment is the only way out.” The research team re-evaluated the source data and determined that these were distinct but complementary dimensions. Consequently, these were synthesized into the comprehensive theme of “financial burden relief.”

Themes were ultimately classified as convergent if they embodied substantively equivalent ideas across methods. An independent reviewer conducted the classification, a second investigator validated all classifications, and discrepancies were resolved through consensus within the research team.

Most participants indicated that their decision to join a clinical trial was influenced by financial considerations, aiming to alleviate the ongoing economic burden of high treatment costs for their families:

Some interviewees, despite having sufficient family resources, expressed a desire to preserve them for the next generation, influenced by traditional family ethics emphasizing “intergenerational sacrifice” and “intergenerational transfer of resources.”

Concerns regarding potential adverse drug reactions were noted, with some individuals expressing fears about their ability to tolerate such effects:

Findings revealed that doubts about therapeutic efficacy and safety risks contributed to participants’ hesitancy toward clinical trials. Refusal to take unapproved research drugs was often attributed to perceived safety issues and limited technical maturity:

Some participants perceived that the clinical data provided to them prior to enrollment were insufficient for informed decision-making, particularly regarding key outcome probabilities such as survival, recovery, and mortality rates:

Early clinical trial risks emerged as a notable concern. Participants expressed greater apprehension toward phase I trials due to unclear side effects, whereas phase III trials with more comprehensive safety and efficacy data were generally viewed as more acceptable:

Uncertainty about the randomization process was identified as a concern during the trial:

Several participants expressed that their limited understanding of clinical trials directly contributed to feelings of uncertainty and apprehension regarding participation:

Some patients develop misconceptions due to a lack of understanding of clinical trials:

Several participants recognized that clinical trials provide not only individual treatment opportunities but also hold forward-looking value for the medical field. They emphasized that the effectiveness of real-world data is critical for breakthroughs in disease treatment and for broader progress in health care:

Participants frequently cited trust in their physicians as a key factor in their willingness to participate in clinical trials. The reputation and recommendation of the treating physician provided a sense of reassurance and credibility.

A subset of participants reported positive experiences with the medical team during the clinical trial, noting that the services provided were excellent and that procedures were flexibly adapted to meet their needs:

Participants reported that the frequency of follow-up visits created a significant time burden and physical fatigue, causing practical disruptions to their daily lives:

Most patients reported no prior exposure to clinical trials, leading to fear rooted in unfamiliarity. Once diagnosed, they faced a lack of formal channels for acquiring reliable information:

Due to the specialized nature of clinical trials, these medical resources may be inaccessible in some remote regions. As a result, professional knowledge regarding clinical trials is largely concentrated in major cities, leading patients in outlying areas to be wary of and avoid them:

The majority of respondents enrolled in clinical trials only after their existing treatment failed, leaving them with no alternative options. This suggests that patients believe clinical trials are their only hope, consistent with previous research results [38,39]. Therefore, patients’ enrollment decisions are strongly shaped by therapeutic necessity and the perceived opportunity to access otherwise unavailable interventions [40].

For some patients, the hope and persistent belief that they can be cured of their illness help them regain some normalcy in their lives and enable them to cope with the severe prognosis of their condition [41]. Participants frequently described trial enrollment as a last resort, underscoring the critical role of perceived therapeutic benefit in decision-making. The prevalence of therapeutic necessity necessitates ethically nuanced communication. Recruitment materials should clearly delineate research objectives from therapeutic possibilities, acknowledge the experimental nature of interventions, and avoid implicit guarantees of clinical benefits. Risk-benefit discussions should balance empathy with transparency, clearly distinguishing between research aims and potential clinical benefits. To retain these motivated yet vulnerable participants, proactive emotional support, dynamic consent processes, and ongoing individual education are recommended.

The physician-patient relationship is a critical determinant of patients’ willingness to participate in clinical trials. Consistent with prior research, strong trust in a physician correlates with greater willingness to participate [38], whereas institutional skepticism tends to heighten patient concerns [42]. However, recent qualitative findings by Murphy et al [43] reveal that such trust can be ethically ambivalent. While it facilitates engagement, it may also compromise informed consent by leading patients to agree to participate out of a desire to please their physicians or to maintain a valued relationship. This underscores the need for a more balanced communication framework that respects emotional dynamics while maintaining ethical integrity. From the physician’s standpoint, this entails moving from the role of educator to that of collaborator, who acknowledges patients’ emotions, encourages open dialogue, and ensures patients feel heard and understood [44]. Such an approach not only fosters authentic trust but also strengthens shared decision-making and mitigates the risk of undue influence. At the same time, it is emphasized that communication strategies must prioritize verifying patient autonomy over merely relying on trust to increase enrollment.

The findings affirm that economic considerations are a powerful component of the decision-making calculus for many patients. This is consistent with studies such as Sun et al [45], which rank free treatment as a significant motivator. However, this economic driver creates a critical paradox when contrasted with research by Kim et al [46], which documents that socioeconomically disadvantaged populations—the very groups expected to be most responsive to financial incentives—are under-represented in US trials. The simplistic explanation of “cultural differences” is insufficient. Instead, this divergence powerfully illustrates that economic incentives are a necessary but not sufficient condition for participation. Our findings, situated in the Chinese context, highlight the potent “pull” of financial relief. Kim et al [46] work, in turn, reveals the formidable structural barriers, such as low health literacy, logistical hurdles, and historical mistrust, that can completely neutralize this pull for vulnerable populations. This synthesis suggests that the efficacy of an economic incentive is profoundly mediated by a patient’s social and structural context. Therefore, for recruitment strategies to be truly equitable, they must move beyond merely offering financial relief and actively dismantle the systemic barriers that prevent vulnerable patients from even considering the opportunity.

From the perspective of the service process, participants’ willingness to engage in clinical research is often undermined by procedural burdens such as repeated examinations, long waiting times, and frequent sampling. These requirements cause physical fatigue and psychological disengagement, ultimately reducing retention. Prior research has sought to mitigate this issue through logistical optimization, such as streamlining online follow-ups, improving communication interfaces, and implementing virtual trial workflows, digital tools, and wearables [47,48]. These measures often represent incremental refinements to existing site-centered models rather than structural solutions to patient burden. In contrast, emerging approaches such as patient-centric sampling and decentralized clinical trials aim to address the root causes of these burdens by integrating research into daily life [49,50]. However, the key insight from the data is not simply that technology is superior, but that minimizing physical disruption to the patient’s routine is essential for retention. Therefore, efforts to improve participation should prioritize reducing the protocol’s physical and logistical burden, rather than merely digitizing communication within a burdensome framework.

Consistent with previous studies, participants’ willingness to participate was influenced by perceived trial risks [23]. The concerns expressed centered on potential adverse reactions, arrangements for posttrial care, and, most critically, the trial’s efficacy. This highlights the urgent need for effective communication to explain potential benefits and risks and to address patient concerns [40]. However, this assessment of risk is not a fixed calculation. Emerging evidence indicates that it is a dynamic process, powerfully shaped by the patient’s evolving clinical reality rather than by the mere duration of their illness. Risk tolerance, for example, often increases with disease severity or after other treatments have failed, as the focus shifts toward potential efficacy. Conversely, a personal history of adverse effects can lead to significant risk aversion [51]. This implies that a one-time, upfront disclosure of risks is insufficient. Instead, risk communication should be regarded as a continuous, adaptive process that evolves alongside the patient’s clinical condition and emotional state. Proactive monitoring of patients’ perceptions and regular follow-up discussions can help identify shifting concerns, clarify misunderstandings, and reinforce trust in the research team. Integrating this dynamic communication model into participant management and recruitment frameworks not only enhances ethical compliance but also strengthens patient retention and data reliability, thereby contributing to the overall integrity and success of clinical trials.

Consistent with previous research, the randomized design of clinical trials remains a major barrier to patient participation [52]. Patients prefer that treatment plans be determined by their physicians rather than randomly assigned [24]. Murphy et al [43] noted that this resistance is often compounded by misunderstandings in which patients erroneously believe that physicians control the allocation. However, Wegge-Larsen et al [5] found that, even when patients fully understood randomization, they remained willing to participate to access novel interventions. This contrast suggests that the barrier of randomization is not absolute but context-dependent. For patients with limited therapeutic options, the opportunity to access a new treatment outweighs the uncertainty of randomization. Conversely, when established treatments are available, randomization is more likely to be perceived as a loss of care. Thus, the effective communication of randomization requires addressing valid patient concerns about care equity, rather than treating resistance solely as a lack of comprehension.

This study’s identification of autonomous, family-delegated, and joint decision-making patterns reveals the complex locus of agency in clinical trial participation. These findings reflect not only individual preferences but also the deep-seated cultural and social contexts within which medical decisions are made. While autonomous decision-making reflects a growing awareness of self-determination, the prevalence of family involvement underscores the influence of values that prioritize collective judgment. This observation aligns with broader findings in oncology care. Systematic reviews by Dijkman et al [21] and Tielemans et al [53] confirm that family involvement in treatment decision-making is a widespread phenomenon, observed across North American, European, and Asian contexts, challenging the notion that it is an exception to the rule of individual autonomy. The regulatory mandate for purely independent decision-making [54], while intended to protect the patient, creates significant tension with this reality. Our findings align with prior research showing that for vulnerable patients, including older adults and those from disadvantaged backgrounds, the ideal of pure autonomy is often practically unattainable, with many ceding decision-making to health care professionals [40]. The role of the family further complicates this picture. Research by Kogan et al [22] powerfully illustrates this dynamic, revealing that caregivers in phase 1 trials often feel compelled to silently support the patient’s choice, withholding their own opinions to protect the patient’s sense of agency. This act of “silent support” reveals a critical nuance: family involvement is not always about co-opting the decision, but is often a mechanism for endorsing and enabling patient autonomy within a supportive relational context. Therefore, the direct transplantation of a universal, Western-centric informed consent model is not merely inappropriate; it is misaligned with patients’ and their families’ lived experiences. It fails to recognize what the evidence consistently shows: the decision-making unit is often the patient-family dyad, not the patient in isolation. A more refined understanding of autonomy is needed, one that transcends rigid individualism and embraces a model of relational autonomy. This perspective preserves the ethical foundation of informed consent while acknowledging the meaningful role of family in decision-making. It recognizes autonomy as contextually shaped, validating both independent and family-informed choices as legitimate expressions of self-determination.