The search strategy consisted of 4 concept blocks (cancer, AI, DHT, and HRQOL) and combined free-text and thesaurus terms using AND/OR Boolean operators. Initial terms were identified by the first author (AI) through scoping searches and key articles in the field. The strategy was refined iteratively in consultation with a co-author (HCL) to capture relevant synonyms, acronyms, and spelling variants, and it was reviewed by an experienced librarian. Search strings were adapted for each database, and we focused on articles containing any of the terms in the title, abstract, and, where possible, the keywords. No further filters or language restrictions were applied at the search stage.

We conducted an initial search on May 20, 2025, in PubMed, PsycINFO, Embase, Scopus, CINAHL, and the Cochrane Library, covering records published from January 1, 2020, to the search date. The start date (January 1, 2020) was chosen to account for the rapid development of AI and to ensure the inclusion of the most recent evidence [20,21]. A second search was conducted in the same databases on August 26, 2025, adding 3 terms that had not been included previously. Therefore, we obtained a large number of duplicates. The reference lists of all included publications were screened to identify additional potentially relevant studies.

Reports were eligible if they (1) enrolled cancer survivors at any disease stage, including those undergoing active treatment or palliative care (in accordance with the definition of the National Cancer Institute) [22], (2) evaluated patient-facing DHTs that explicitly included AI components (AI-enabled DHTs), and (3) reported an assessment of HRQOL. For the purposes of this review, AI-enabled DHTs refer to patient-facing DHTs that incorporate an AI component, whether embedded within the platform or operating as an external module.

Nonoriginal reports, including other reviews, editorials, commentaries, conference abstracts, and study protocols, were excluded. Only studies that included a control group (eg, usual care and non–AI-enabled DHTs) were eligible for inclusion in the meta-analysis.

Although we initially considered including non-English articles relying on software-generated translations, these were ultimately excluded because we could not validate them with a native speaker and therefore could not avoid potential misinterpretation. However, these studies are cited in the results for completeness.

All identified references were screened using Covidence (Veritas Health Innovation Ltd) [23], and duplicates were automatically excluded. Two authors (AI and PC or KETT) independently screened all titles and abstracts against previously agreed-on inclusion and exclusion criteria. The same 2 authors screened the full texts of all articles previously included by at least one author to make a final decision on inclusion. Any disagreements were resolved through discussion or by consulting a third author. The study selection process is summarized in a PRISMA 2020 flow diagram.

Data extraction was supported using SciSpace [24] and Perplexity [25]. All machine-extracted information was verified against the original manuscripts to ensure that no relevant information was misinterpreted or missed (AI) and validated by a second author (PC or KETT). Data were extracted from the published articles and any supplementary materials, and no individual participant data were requested. Where information was unclear or insufficiently detailed, we contacted the authors for clarification. However, no responses were received.

Collected data included (1) study information: title, authors, year, journal, country, and study design; (2) sample characteristics: sample size, sociodemographic information, and cancer-related characteristics for both intervention and control groups; (3) DHT details: type, purpose, AI components, input data sources, and output type; (4) HRQOL outcomes: baseline and postintervention measurements, instruments used, domains assessed, and main findings; and (5) additional PROs: instruments used and key results. When multiple articles reported on the same study, all relevant publications were considered to provide a comprehensive description of the methods and outcomes.

For randomized controlled trials (RCTs), the risk of bias was assessed using the revised Cochrane Risk of Bias 2 (RoB 2) tool [26], and for non-RCTs, the ROBINS-I V2 (Risk of Bias in Non-Randomized Studies—of Interventions, Version 2) [27]. Single-arm studies were not formally assessed for risk of bias and were reported narratively without inclusion in any quantitative synthesis.

Two authors (AI and LKJ) independently evaluated each study across all domains, with possible ratings of low risk, some concerns, or high risk (RoB 2), or low, moderate, serious, and critical concerns (ROBINS-I V2). The evaluations were then compared, and any discrepancies were resolved through discussion. When multiple articles reported on the same study, all relevant publications were considered to provide a comprehensive assessment.

For the purposes of this systematic review, studies were grouped as controlled (including randomized and nonrandomized trials) or single-arm studies. Controlled studies were eligible for quantitative synthesis (meta-analysis), whereas single-arm studies were only summarized narratively.

A comprehensive search of 6 scientific databases identified 2340 records (Figure 1). After removing duplicates (n=1417), 923 records were screened. Of these, 30 full-text articles were assessed for eligibility, and 8 publications from 7 studies (N=2867 participants) met the inclusion criteria and were included in the review.

Two potentially relevant records published in Chinese were excluded because reliable translation and validation were not feasible. One was a pilot study from an already included study [30], and the other reported data from a distinct study not otherwise included in this review [31]. Citation searching did not identify any additional eligible studies.

The 7 included studies comprised 2 single-arm studies [32,33], 3 RCTs [34-37], 1 partial-crossover RCT [38], and 1 non-RCT [39] (Table 1).

The studies covered multiple cancer types and survivorship stages (Table 1). RCTs by Schmitz et al [38] in the United States and by Hu et al [34] and Jiang et al [35] in China focused exclusively on breast cancer survivors, while a US single-arm study by Horesh et al [33] included both breast and ovarian cancer survivors. Together, these studies targeted survivorship stages from active treatment through early survivorship to metastatic disease. The RCT by Li et al [37] examined early-stage postoperative lung cancer survivors in China, and the RCT by Kamdar et al [36] focused on patients with metastatic solid tumors receiving palliative care in the United States. The remaining studies included survivors of any cancer type in different contexts. The non-RCT by Zhang et al [39] focused on patients receiving chemotherapy in China, and the single-arm study by Buchan et al [32] included individuals at any stage of disease in the United States. Quality of life was assessed in all included studies as a primary or secondary outcome, and most studies used validated HRQOL instruments.

Of the 4 RCTs, 2 were rated as low risk of bias, and 2 studies were assessed as having some concerns. However, all concerns were observed in different domains and did not indicate a consistent source of bias across the evidence base. The non-RCT by Zhang et al [39] was found to be at serious risk of bias, primarily due to inadequate control of confounding factors in a time-separated convenience sample. There were additional concerns regarding the unclear management of exclusions and missing data. Furthermore, inconsistencies in the reporting of results suggest caution when interpreting the findings. No studies were evaluated as high (RoB 2) or critical (ROBINS-I V2) risk of bias. Therefore, none were excluded. Further details on the risk of bias evaluation are available in Figure 2.

Multimedia Appendix 2 presents the magnitude and direction of changes for all presented outcomes.

In addition to HRQOL, the included studies reported other PROs, including user satisfaction and adherence, mental health, physical symptoms, additional PROs, and adverse effects.

This systematic review synthesized evidence on available patient-facing AI-enabled DHTs and their effect on HRQOL and broader PROs in cancer populations. Eight publications from 7 studies conducted in China and the United States evaluated a range of mobile- and web-based interventions, most of which aimed to support at-home symptom management. AI was used to personalize recommendations (eg, exercise, nutrition, and symptom triage), adapt delivery formats (eg, VR and voice interfaces), and tailor educational content. Overall, the effects on HRQOL were mixed. Results from the narrative synthesis showed that among the 5 controlled studies, all 3 RCTs conducted in China found significant improvements across multiple HRQOL domains, including global health status, role, cognitive, emotional, social, and functional well-being, while neither of the 2 US RCTs found significant between-group effects. The 2 single-arm US studies suggested within-group improvements, although the absence of control groups limits the interpretability of those findings. An exploratory meta-analysis of 5 HRQOL domains showed modest improvements only in general health and social functioning, but no effect on physical functioning, role functioning, and emotional well-being. However, these results should be interpreted with caution given the small number of contributing studies, limited sample sizes, heterogeneity, and the presence of risk-of-bias concerns in individual studies. Other PROs, including mental health, physical symptoms, self-efficacy, and social support, also showed mixed but generally encouraging results. Mental health outcomes were variable, with some DHTs reducing distress or anxiety and others showing no effect or small, but not clinically significant, increases in anxiety. Despite considerable heterogeneity in both interventions and populations, most DHTs were associated with improvements in key physical symptoms, including pain, fatigue, and sleep. User adoption and satisfaction were generally high, and only mild adverse effects were reported occasionally.

Several patterns might help explain the variation in effects across studies and guide the design of future AI-enabled DHTs for cancer survivorship care. Improvements in HRQOL and other PROs were more frequent in the trials conducted in China, particularly for general health, whereas effects were less consistent across specific HRQOL domains. Two trials conducted in the United States showed no benefits in HRQOL, while 2 observational studies from the United States lacked control groups. Methodological factors, including small samples, heterogeneous measures, and short follow-up periods, likely contributed to these differences. In digital oncology research, changes in HRQOL might require sustained use of DHTs over time, as physical and psychological symptoms develop gradually [40,41]. Trials with short follow-up periods may, therefore, have lacked sufficient time to detect change, even when adherence was high. Another explanation might be differences in perceptions of AI-enabled DHTs and in-person consultations across cultures [42]. These may influence how patients interpret symptom feedback, interact with automated recommendations, and integrate digital tools into their routines. Furthermore, cross-national research suggests that social desirability bias varies across cultures, with relatively higher levels observed in China [43]. This may have led to more favorable evaluations of AI-enabled DHTs, especially in clinician-administered surveys. However, these cultural explanations remain speculative and cannot be distinguished from other country-level differences.

Our findings should be interpreted in the context of heterogeneous definitions of cancer survivorship across studies, which translate into substantial differences in symptom burden and care needs. While some interventions targeted patients with metastatic or advanced disease receiving ongoing or palliative treatment, others focused on disease-free survivors after curative therapy. Consistent with this heterogeneity, trials reporting no improvements in HRQOL tended to include patients with metastatic disease. Previous studies have shown that, among patients with high symptom burden in palliative care settings, supportive DHTs alone are often insufficient to improve overall HRQOL, although they may alleviate specific physical symptoms [44]. Our results have partially confirmed these findings. While AI-enabled tools may offer more personalized and proactive support than traditional DHTs, their benefits may be limited in advanced disease settings. This suggests that AI-enabled interventions should be integrated with comprehensive supportive care to improve HRQOL in advanced cancer.

Studies showing positive effects on HRQOL and PROs often included human involvement. This aligns with broader evidence that digital tools can initiate but rarely sustain long-term engagement among populations with complex needs without human involvement [45]. Hybrid models that combine AI-enabled monitoring and personalization with clinician or nurse follow-up leverage the strengths of both technology and human support. Thus, they can better translate digital engagement into behavior change and symptom improvement. Interventions that supported question-asking and peer discussion also reported positive effects on PROs. Features that enable communication with clinicians or connect patients with peers may increase self-efficacy, normalize experiences, and reduce isolation—mechanisms that are widely recognized in cancer survivorship research [46]. However, AI components themselves may introduce bias. For example, NLP or large language model–based systems that classify patient input or triage symptoms can introduce systematic measurement shifts, which could affect clinical decision support [47,48]. These findings highlight the value of designing comprehensive DHTs as platforms for holistic support.

User satisfaction and engagement were high across studies, indicating strong acceptability of DHTs containing AI components in cancer survivorship care. Notably, the effects observed on HRQOL and other PROs are similar to those reported for previous digital health interventions without AI [41,46]. However, because none of the studies included equivalent DHTs without AI components, we cannot determine whether these effects are attributable to the AI functionality, other DHT features, or nonspecific effects. Therefore, integrating AI might not significantly alter patients’ experiences, either by enhancing outcomes or introducing new barriers. Instead, the effectiveness of DHTs on PROs seems to depend primarily on design quality and relevance of support rather than on whether the underlying system relies on AI. Future research should focus on larger, multisite, longitudinal trials, more diverse populations with cancer, the use of traditional DHTs without AI components for control purposes, and the identification of different engagement trajectories among users. In practice, AI-enabled DHTs may support HRQOL when paired with adequate human follow-up, peer-support features, and careful patient selection based on disease stage, symptom burden, and digital readiness.

Our findings should be interpreted in light of several limitations. First, clinical heterogeneity limits the validity of pooled estimates. The included studies covered diverse types of cancer and stages of survivorship, making a single pooled estimate for cancer survivors problematic. Furthermore, the 3 different HRQOL instruments used across studies introduce measurement heterogeneity. Readers should, therefore, interpret findings from the exploratory meta-analyses with caution.

Second, the small number of eligible studies and participants limits the statistical power of the meta-analyses. While the narrative synthesis provides additional insights, the results from the meta-analyses should be considered exploratory.

Third, methodological limitations across studies reduce comparability and generalizability and may inflate observed effects. These limitations include small sample sizes, lack of control groups or randomization, short follow-up periods, and reporting limitations. None of the studies included equivalent DHTs without AI components, so it is not possible to determine whether the observed benefits are due to AI functionality, other DHT features, or nonspecific effects.

Fourth, geographic confounding limits generalizability. The limited geographical diversity (primarily studies from China vs US) prevents the isolation of intervention effects from country-level factors such as health care systems, cultural context, or implementation characteristics.

Fifth, how AI was defined and operationalized varied across studies, ranging from rule-based decision support to ML algorithms. This complicates direct comparisons across interventions and limits the interpretability of pooled findings. Furthermore, although the search strategy was intentionally broad and included multiple AI-related terms, studies that did not adequately describe the technical components of their DHTs may have been missed. Future reviews would benefit from applying a prespecified AI classification framework.

Finally, research on AI-enabled DHT is advancing rapidly. The numerous protocols identified during the screening process suggest that more publications are forthcoming. A living systematic review approach may, therefore, be valuable for maintaining an up-to-date evidence base.

This review provides a timely synthesis of evidence on AI-enabled DHTs for cancer care and establishes a conceptual and methodological foundation for future work. Although effects on HRQOL were mixed, some interventions showed benefits across HRQOL domains and other PROs. Importantly, as none of the studies used control groups with equivalent DHTs without AI components, the observed benefits cannot be attributed specifically to the AI components. Our findings are preliminary, and it is too early to conclude whether patient-facing AI-enabled DHTs improve cancer survivorship care. Future research should employ active control conditions to isolate the contribution of AI features specifically.