Cancer remains a leading cause of morbidity and mortality worldwide, placing a substantial burden on health care systems. In 2022, an estimated 20 million new cancer cases and 9.7 million cancer-related deaths occurred worldwide, with approximately 53.5 million individuals living with cancer [1]. A cancer diagnosis not only presents significant challenges to health care systems but also profoundly affects patients’ daily lives. Both cancer itself and its treatment can result in a wide range of physical symptoms and psychological distress [2]. These multifaceted burdens frequently disrupt patients’ overall well-being and significantly impair their quality of life [2].

In this context, there has been increasing emphasis on developing effective cancer care that provides continuous, comprehensive support alongside standard oncological treatment, aiming to address patients’ physical and psychological needs and to promote behavioral change across the cancer care continuum [3]. However, implementing such care models through traditional face-to-face delivery remains difficult, as these approaches are often constrained by time, staffing, and scalability—particularly in resource-limited settings [4]. As the demand for personalized and continuous support grows and health care systems shift toward cost-efficient care, these limitations have become increasingly evident, highlighting the necessity for automated, scalable, and cost-effective solutions to enhance care for patients with cancer [5].

In response to these limitations, there has been growing interest in leveraging mobile health (mHealth) technologies to enhance care delivery. In particular, conversational agents (CAs) have emerged as a promising mHealth innovation in supporting care for patients with cancer [6]. CAs are dialogue systems capable of understanding and generating human language to enable effective human-computer interactions [7]. Unlike conventional mHealth tools—which often rely on static content, unidirectional reminders, or resource-intensive human support—CAs provide automated, real-time dialogue capabilities that support dynamic, personalized interactions [8]. By simulating human-like conversations, CAs can adapt to patients’ evolving informational, emotional, and practical needs—an essential feature in the complex and often unpredictable trajectory of cancer care. These capabilities may not only improve the continuity and personalization of care but also offer a scalable means of support, potentially reducing the burden on health care providers [9].

Although CA interventions have shown promise and are increasingly used in cancer care, evidence regarding their feasibility and effectiveness remains inconclusive [10]. Although some studies have reported positive outcomes—such as improved symptom management, psychological support, behavior change, and quality of life—others have shown minimal or no effects [11,12]. These inconsistencies highlight the need for a comprehensive synthesis of existing evidence to clarify the effectiveness of CAs in supporting cancer care [10]. In parallel, although feasibility is frequently cited as a strength of CA interventions—due to their ability to deliver personalized support, reduce access barriers, and provide timely responses—several concerns have also emerged, including risks of miscommunication, limited empathetic responsiveness, and challenges in sustaining use [13,14]. Together, these inconsistencies in both clinical and implementation outcomes point to the urgent need for a systematic synthesis to guide the design and integration of CA interventions in cancer care.

To date, no systematic review has comprehensively evaluated the effectiveness or feasibility of CA-based interventions in cancer care. Existing reviews remain limited in scope; notably, only 1 previous narrative review provides a broad overview of CA applications in oncology but does not quantitatively synthesize outcomes or systematically synthesize implementation outcomes [10]. Consequently, it remains unclear whether CAs lead to improvements in health outcomes of patients with cancer; what implementation-related patterns—particularly regarding feasibility, usability, and acceptability—have emerged; and what core features define current CA interventions for cancer care. A rigorous synthesis of the existing evidence is warranted to inform future CA development, facilitate clinical implementation, and support evidence-based decision-making.

This systematic review and meta-analysis aimed to comprehensively evaluate the use of CA interventions in cancer care. Specifically, the objectives were to (1) assess their effectiveness in improving patient outcomes, including quality of life, pain, anxiety, depression, psychological distress, physical activity, treatment-related side effects, and health information acquisition; (2) evaluate implementation outcomes—such as feasibility, acceptability, and usability—to determine the practical applicability of CAs in oncology settings; and (3) synthesize the core characteristics of CA interventions, including intervention components, and CA delivery and interaction logic, as well as input and output modalities, to inform future development and implementation.

This systematic review and meta-analysis was based on a prespecified protocol registered with PROSPERO (International Prospective Register of Systematic Reviews; CRD42025645982) and conducted and reported according to the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) statement [15]. The PRISMA checklist can be found in in Table S1 in Multimedia Appendix 1.

A comprehensive systematic literature search was conducted on 4 different electronic databases—PubMed, Embase, Web of Science, and Cochrane Library—from the index date through February 3, 2025. After study screening and selection, references for the identified studies and relevant websites, such as clinical trial registries, were manually searched to find potential related gray papers. The keywords of the search mainly revolved around “cancer” and “conversational agent.”. The full search strategies and the number of search results from each database are detailed in Table S2 in Multimedia Appendix 1.

We considered eligible randomized controlled trials (RCTs) and nonrandomized interventions (NRIs) that examined the effects of CAs, defined as dialogue systems capable of understanding and generating human language to enable effective human-computer interactions. Eligible studies aimed to improve health-related outcomes in adult patients with cancer using CAs that were either specifically developed for oncology care or originally designed for general health purposes and subsequently applied to patients with cancer. The inclusion and exclusion criteria are reported in Textbox 1.

All studies identified through database searches, reference lists, and relevant websites were imported into EndNote X9 (Clarivate Analytics), a reference management software program. EndNote was also used to detect and remove any duplicate entries. The selection of studies was conducted in 2 stages. Initially, 2 independent reviewers (authors JXH and YXH) screened the titles and abstracts of all identified studies in EndNote to exclude studies that did not meet the basic inclusion criteria. Subsequently, the full texts of the remaining studies were retrieved and assessed in detail against the inclusion and exclusion criteria by the same 2 reviewers. Any disagreements between reviewers were resolved through discussion or, if necessary, consultation with a third reviewer (author PJS).

Data extraction using Microsoft Excel was first pilot-tested on 3 studies and revised with additional headings before being applied to all included studies. Two independent reviewers (JXH and YXH) conducted the data extraction process, with any disagreements resolved by a third reviewer (PJS). The following data were extracted: (1) basic study characteristics (author, title, country, and publication year), (2) sample characteristics (cancer type, sample size, age, and gender), (3) intervention components (consultable content, CA-only vs multicomponent interventions, frequency and duration, theoretical framework), (4) CA delivery and interaction logic (CA name, delivery platform, prompt sequence, whether the CA was cancer specific, clinician referral, algorithm type, and embodiment type), (5) input and output information (input data and format, input and output modality, and output source), and (6) outcome measures (health outcomes, feasibility, acceptability, and usability).

Two independent authors (JXH and YXH) assessed the risk of bias for included studies using the revised Cochrane Risk of Bias Tool version 2.0 (ROB 2.0) for RCTs [16] and the Joanna Briggs Institute (JBI) Critical Appraisal Checklist for NRIs [17]. When disagreements arose, a third reviewer (PJS) was consulted, and consensus was achieved through discussion.

The primary outcome was quality of life, selected for its direct relevance in assessing the overall effectiveness of CA interventions in cancer care. Secondary outcomes included effectiveness outcomes (physical activity, pain, anxiety, depression, psychological distress, treatment-related side effects, and health information acquisition) and implementation outcomes (feasibility, acceptability, and usability). Depending on data availability and consistency, either a meta-analysis or a narrative synthesis was conducted.

Effectiveness outcomes were synthesized using meta-analysis when the included studies demonstrated sufficient statistical homogeneity. Analyses were performed using Review Manager (RevMan) version 5.4 (Cochrane Collaboration). For each outcome, either change-from-baseline scores or postintervention scores were extracted, with a consistent approach was applied within each analysis. When necessary, missing summary statistics were estimated following the methods recommended in the Cochrane Handbook for Systematic Reviews of Interventions [18]. Effect sizes were reported as standardized mean differences (SMDs) with 95% CIs when different measurement scales were used, and as mean differences (MDs) when the same scale was applied. Heterogeneity was assessed using the chi-square test and the I² statistic. If P>.10 and I² < 50%, it suggested that the heterogeneity among studies as small, and the fixed-effect model was selected. If P<.10 and I²> 50%, it suggested that there was obvious heterogeneity among the studies, and the random-effects model was adopted [19]. Sensitivity analysis was performed by comparing the results of the combined effect sizes of the fixed-effects model and random-effects model to ensure stability and reliability of the results [18]. Publication bias was examined using funnel plots.

For effectiveness outcomes not amenable to meta-analysis due to insufficient data or heterogeneity, a narrative synthesis was conducted. This synthesis was structured by outcome domain and described the direction and magnitude of reported effects. For implementation outcomes, a narrative synthesis was conducted and organized thematically according to key evaluation dimensions, including feasibility (retention rate, safety, interactions, duration of engagement, and engagement rate), acceptability (satisfaction, perceived helpfulness, and recommendation willingness), and usability (usability of content, ease of use, and user experience). In addition, intervention characteristics—including intervention components, CA delivery and interaction logic, and input and output information—were descriptively summarized.

The database search yielded 2087 results. A total of 1323 (63.4%) titles and abstracts were screened after removal of duplicates, and 1188 (56.9%) studies were excluded after screening. In total, 135 (6.5%) potential full-text papers were identified for further evaluation, of which 15 (11.1%) met our inclusion criteria. An additional 87 papers were identified from websites and relevant citations. After removing duplicates and assessing titles, abstracts, and full texts, 2 extra papers were identified and included. In total, 17 papers were included in this systematic review (Figure 1).

The 17 included studies [6,11,12,20-34] were published between 2016 and 2024 (please note that Bema [27] contains the complete documentation of results from Mlakar et al’s [26] study). There were a total of 1817 participants, and sample sizes ranged between 30 (1.7%) and 336 (18.5%). The mean age of the participants ranged between 25 (SD 2.9) and 64 (SD 11.8) years. Of the 17 studies, 12 (70.6%) were RCTs, and 5 (29.4%) were NRIs. Most studies were conducted in the United States (n=10, 58.8%) and primarily targeted patients with breast cancer (n=8, 47.1%). Overall, 69.7% (n=1266) of the participants were women. A summary of study characteristics is presented in Table 1.

The characteristics of the CA interventions are presented in Table S3 in Multimedia Appendix 1.

The studies included in the qualitative synthesis of feasibility, acceptability, and usability are listed in Table S4 in Multimedia Appendix 1.

Of the 17 studies, 10 (58.8%) were included in a meta-analysis, comprising 6 (60%) RCTs [12,20,22-24,30] and 4 (40%) NRIs [25,26,33,34]. Given the clinical and methodological similarities across the included studies, quality of life, physical activity, pain, anxiety, depression, and psychological distress were adopted as outcome indicators of the meta-analysis. The results of the meta-analysis are presented in Table 2 and Figure 2. Sensitivity analyses comparing fixed- and random-effects models for each of these outcomes yielded consistent results, supporting the stability of the pooled effect estimates (see Figures S1-S6 in Multimedia Appendix 1).

The results of risk-of-bias assessments of the 17 studies are reported in Figures S7 and S8 and Table S5 in Multimedia Appendix 1. Of the 12 (70.6%) RCTs, 3 (25%) had an overall low risk of bias [21-23], 4 (33.3%) exhibited a high risk of bias [6,12,20,32], and another 5 (41.7%) raised some concerns [11,24,28-30]. Common methodological issues included inadequate randomization, deviations from intended interventions, and bias due to missing outcome data or outcome measurement. In addition, 5 (41.7%) quasi-experimental studies met 7-8 out of 9 quality assessment criteria, indicating relatively good methodological quality [25,26,31,33,34]. Common limitations among these studies included the absence of control groups, limited follow-up reporting, and insufficient pre- and postintervention outcome measurement.

This systematic review is the first to comprehensively evaluate the effectiveness and feasibility of CA interventions in supporting care for patients with cancer. In total, 17 trials involving 1817 participants were included, comprising 12 (70.6%) RCTs and 5 (29.4%) NRIs. Of these, 10 (58.8%) studies were included in a meta-analysis. The meta-analysis findings indicated that CA interventions are effective in promoting physical activity, reducing pain and anxiety, and improving the quality of life among patients with cancer. However, no significant effects were observed for depression or psychological distress. Narrative synthesis suggested that CAs have the potential to enhance health information acquisition and help manage treatment-related side effects. Furthermore, CAs were generally found to be feasible, acceptable, and usable among patients with cancer, particularly during the initial phase of use. However, user engagement tended to decline over time, underscoring the need for strategies to sustain long-term use. These results support the potential integration of CAs into broader cancer care frameworks and inform future direction for optimizing CA design to enhance sustained engagement and clinical effectiveness.

Prompt sequencing and interaction modes are central to the design and delivery of CA interventions, shaping how users engage with the CA system and receive tailored support. Across the included studies, CAs adopted user-initiated inputs [11,22,24,28], system-initiated prompts [12,21,26,32], or a combination of both approaches [6,20,23,25,29-31,33,34]. The sequencing of prompts generally aligned with the CA’s intended function and interaction style: user-initiated prompts supported on-demand health queries or symptom check-ins; system-initiated prompts provided scheduled reminders or symptom-triggered alerts; and hybrid models combined both to enhance adherence and engagement [8]. Upon receiving input, CAs typically conducted a needs assessment and generated personalized responses via rule-based scripts [11,12,21,23,33], AI-driven dialogue generation [20,22,26,28,34], or both approaches [6,24,25,29-32], depending on the design and operational mechanisms of the system. The response content of CAs was largely grounded in authoritative sources, including clinical guidelines, health websites, expert consensus, peer-reviewed literature, and evidence-based QnA libraries developed by study teams. Cloud-based platforms were also integrated in some studies to deliver responses from predefined knowledge bases, improving response consistency and system scalability [6,11,20,22,24,34]. Handover mechanisms were incorporated in only one-third of interventions, enabling escalation to clinical staff in high-risk situations to ensure safety and responsiveness [20,23,24,30,31,34]. Additionally, most interventions adopted mobile- or web-based interfaces, though only a third incorporated visual elements (eg, icons, avatars) [20,22,25,26,32].

This review supports the effectiveness of CAs in cancer care, particularly in promoting physical activity and alleviating symptoms, such as pain, treatment side effects, and anxiety, thereby enhancing patients’ overall quality of life. Although prior studies in chronic disease populations have also reported favorable outcomes with CAs, many of those reviews focused on relatively narrow objectives—such as physical activity or weight management—with limited attention to multidimensional outcomes [35,36]. In contrast, the cancer-specific CAs included in this review demonstrated broader functionality, addressing not only physical activity but also a wide range of symptoms and overall quality of life. This broader scope and effectiveness may be attributed to the interactive and tailored content of CAs, which aligns well with the complex and evolving needs of individuals undergoing cancer treatment [37]. Specifically, reductions in the symptom burden—such as pain, anxiety, and treatment-related side effects—highlight the potential of CAs to deliver symptom-specific education and practical self-care strategies [7]. Improvements in physical activity observed across studies may reflect the effectiveness of CA features, such as timely reminders, individualized feedback, and structured goal setting [38]. These benefits are particularly important in oncology, where preserving physical function and controlling symptoms are critical to maintaining quality of life [38]. Collectively, these features position CAs as a promising adjunct to conventional oncology care, with the potential to improve multidimensional health-related outcomes and enhance supportive care delivery.

However, the nonsignificant effect of CAs on depression and psychological distress requires further exploration. Both conditions represent complex, multifaceted, and long-lasting psychological challenges that require deeper emotional empathy and theoretically grounded interventions [39]. Although current CA systems may help alleviate transient psychological states, such as anxiety, by delivering multidimensional informational support, they remain inadequate for more complex and persistent psychological demands [37]. This inadequacy is largely attributable to the structural limitations of most existing CAs, which are predominantly rule based or minimally AI enhanced [10]. As a result, they are unable to recognize complex psychological states, respond empathetically, or support deep emotional regulation [40]. Notably, recent studies in other fields suggest that AI-powered CAs based on large language models (LLMs) may even outperform physicians in perceived empathy—possibly due to their advanced attention mechanisms, emotionally rich training data, and context-aware communication design [41]. Although these advances have not yet been applied in cancer care CAs, their integration holds promise. Future CA systems could incorporate emerging technologies, such as AI and LLMs, to enhance empathic communication and provide more comprehensive psychological support [10].

Narrative synthesis revealed that studies using single-session CA interventions primarily reported immediate informational benefits, specifically enhanced health information acquisition. In contrast, sustained health improvements—such as increased physical activity, better symptom management, and enhanced quality of life—were typically observed in studies that implemented multisession or extended-duration interventions lasting from 4 weeks to 6 months. These findings suggest that although brief, single-session interventions may suffice for improving informational outcomes, more sustained exposure—of at least 4 weeks—may be necessary to achieve meaningful improvements in sustained health benefits [8]. However, due to the limited number of studies per outcome and inconsistencies in intervention duration reporting, the optimal intervention duration could not be formally examined. Future research should investigate the relationship between intervention duration and effectiveness to determine the minimum exposure required for achieving sustained benefits across different domains of cancer care [42].

CAs were generally found to be safe, feasible, acceptable, and usable. However, suboptimal long-term engagement remains a critical challenge, underscoring the need for strategies to sustain continued use. This trend is consistent with findings from prior CA interventions in chronic disease management, where high attrition rates were commonly reported despite favorable early-phase implementation outcomes [43,44]. Several design features may have hindered this sustained engagement. Notably, two-thirds of included CAs were fully automated and lacked handover mechanisms to escalate complex or high-risk queries to clinical personnel, potentially weakening users’ sense of accountability and connection, thereby undermining long-term engagement [44]. In addition, only one-third incorporated visual elements (eg, avatars, icons), which may have limited interactivity and reduced the overall appeal of the interface [45]. Moreover, most CAs relied on narrow content databases, leading to repetitive information and diminishing novelty over time [10]. To address these issues, future CA interventions should embed health care professionals into the interaction loop, incorporate emotionally engaging visuals, and leverage advanced technologies—such as LLM-based interfaces and enriched data sources—to provide diverse, adaptive content that evolves with user needs [46]. These enhancements may improve long-term engagement, build trust, and strengthen the effectiveness and scalability of CA-based interventions [44].

This study presented the first systematic review and meta-analysis to comprehensively synthesize the characteristics, feasibility, and effectiveness of CA-based interventions in supporting care for patients with cancer, providing an up-to-date synthesis of the current evidence and inform future clinical practice and intervention development.

However, several limitations should be acknowledged. First, this review included only peer-reviewed studies published in English, which may have introduced publication and language bias (results are presented in Figures S9-S14 in Multimedia Appendix 1). This decision was made to ensure methodological rigor and consistency in reporting quality. However, relevant studies in other languages or nonindexed sources may have been overlooked. Future reviews could address this by incorporating non-English and gray literature to improve inclusiveness. Second, heterogeneity in intervention characteristics, such as intervention duration, technology algorithms, and delivery modality, may have introduced variability that warrants cautious interpretation. Given the limited number of eligible trials, formal subgroup analyses were not feasible; therefore, we interpreted outcome variability narratively based on key intervention features. Future meta-analyses with a larger evidence base may facilitate subgroup analyses to examine their potential moderating effects. Third, most of the included studies had relatively small sample sizes, and some were NRIs, which may limit the generalizability of the findings. This limitation reflects the early, exploratory stage of research on CA interventions in cancer care. To mitigate this, we conducted both meta-analysis and narrative synthesis to comprehensively integrate the current evidence on effectiveness. In addition, a separate narrative synthesis of implementation outcomes was performed to capture early insights into CA delivery and contextual applicability. Future well-designed RCTs with larger sample sizes are warranted to enhance the robustness of the evidence base and support broader clinical implementation.

Overall, CA interventions address persistent service gaps in both clinical and community settings. However, their successful integration into routine care requires systematic consideration of diverse aspects. One key aspect identified in our review is the limited application of theoretical frameworks to guide the design, delivery, and evaluation of CA interventions. Future interventions should be guided by established theoretical frameworks to ensure coherent design, targeted content delivery, and meaningful outcome evaluation [47]. In particular, stage-based models, such as the Transtheoretical Model (TTM), which links intervention strategies and outcomes to users’ readiness for change (eg, precontemplation, preparation, action, maintenance), may offer added practical value [42]. Applying such models could help clarify how CAs influence patient outcomes, address individual needs, and reinforce the scientific foundation of future CA interventions [36].

Second, future CA development should build on existing technologies by incorporating more flexible, adaptive, and emotionally responsive capabilities. Many current systems remain limited by predefined scripts or limited AI functionality, restricting their ability to understand context or convey empathy. Integrating advanced techniques, such as LLMs, may improve natural language understanding and emotional responsiveness, leading to greater personalization and sustained user interaction [48]. Additionally, embedding health care professionals into the interaction loop and incorporating multimodal communication features may further enhance safety, accessibility, user trust, and long-term engagement [7].

Third, the successful application of general-purpose health CAs in oncology care highlights their potential for cross-condition adaptability and broader applicability across health care contexts. Although most CAs included in this review were specifically designed for oncology, several originally developed for general health conditions have also demonstrated effectiveness when adapted for oncology care [12,25,29]. This adaptability indicates that well-designed CAs may be transferable across clinical conditions, suggesting valuable directions for future development. Clinically, repurposing CAs may reduce development costs and expedite deployment in supportive care of different diseases [8]. This also supports the development of CA systems with reusable modules, where disease-specific content can be plugged into a standardized infrastructure [7]. This approach enables continuity of care across comorbid conditions by centralizing supportive functions in a unified CA platform. To support broader implementation, future research should examine how contextual adaptations (eg, population, disease, setting) affect CA effectiveness and user experience [49].

This systematic review is the first to comprehensively evaluate CA interventions in supporting care for patients with cancer, with preliminary evidence supporting their feasibility and effectiveness. However, the limited psychological benefits and suboptimal long-term user engagement indicate the need for further refinement. Future research should adopt theory-driven approaches and explore the integration of emerging technologies to enhance personalization, empathy, and sustained engagement in CA interventions. Moreover, rigorous study designs—particularly large-scale RCTs—are needed to evaluate the effectiveness and implementation processes of CAs in cancer care.