The systematic review and meta-analysis were conducted in compliance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines, and the PRISMA 2020 expanded checklist was used (Checklist 1) [20]. The protocol for this study was registered in the PROSPERO (International Prospective Register of Systematic Reviews); no amendments were made to the search protocol after the initial registration (CRD420251125990).

The specific eligibility requirements of inclusion criteria followed the population, intervention, comparison, outcomes, and study format: (1) patients aged 18 years and above who have been diagnosed with CRC, colon cancer, or rectal cancer; (2) can be at any treatment stage (perioperative period, adjuvant chemotherapy period, maintenance treatment period, or rehabilitation period); (3) no restrictions on tumor stage, whether with stoma or metastasis; (4) intervention was delivered via a web-based remote digital health management system, such as mobile apps and AI. The intervention methods include “guided intervention (ie, feedback from health care providers or researchers during the intervention process)” and “unguided self-management intervention (fully automated feedback)” or a combination of both; (5) the comparison was between the control group that received only the usual care interventions and the intervention group that used the web-based remote digital health management system in addition to the usual care interventions; (6) the outcomes were the impacts of the interventions on overall or at least one type of relevant health-related outcomes (eg, NS and QoL); and (7) the study design was randomized controlled trials (RCTs).

The review excluded (1) studies with unclear diagnostic and efficacy criteria; (2) studies that were cohort studies, review articles, case reports, descriptive studies, opinion articles, gray literature, unpublished literature, or abstracts; (3) studies with incomplete or erroneous data that could not be merged; and (4) studies describing protocols, along with those focusing solely on the interface or internal structure of apps; and studies based on web pages or websites without associated apps.

The literature search was initially designed according to our prospectively registered protocol, which restricted inclusion to English-language studies and searched 4 English databases. In the final manuscript, we adjusted the search strategy to enhance comprehensiveness and reduce language bias, and these modifications have been clearly reported as post hoc adjustments. The literature search for this systematic review was reported in accordance with the PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Literature Search Extension) statement [21]. A systematic search was conducted using subject terms combined with free terms, covering the following databases, including PubMed, Embase, Cochrane, and Web of Science. The search period began from the establishment of each database and lasted until March 2, 2026.

This search strategy for this systematic review was developed by drawing on and optimizing strategies from previous studies, and it did not undergo a specific peer review process. The search parameters were set to include only RCTs and peer-reviewed papers, with no restrictions on publication year or language, and no use of published search filters. To maximize the scope of the search, the snowball method was used for citation retrieval, that is, all the references of the included studies and their citations were checked. For missing outcome data in some studies, the relevant authors were contacted via email to supplement the original data that were not obtained through the initial search. This systematic review only used the above methods to retrieve literature. No supplementary search of research registration databases was conducted, nor was there targeted browsing of printed conference proceedings or government websites. The final search date for all databases was March 2, 2026. The detailed search strategy can be found in Multimedia Appendix 1.

FXJ and SHH separately reviewed the identified papers to reduce potential errors and bias throughout the selection process. Initially, the authors screened the titles and abstracts of the potential papers against the set inclusion and exclusion criteria. Following this, the final selection of papers was made after thoroughly reading the complete manuscripts of the qualifying papers and their references. Any discrepancies were settled through discussions among the authors until a consensus was reached.

A structured data extraction form was used to gather the following details: the first author’s name; year of publication; country; study type; participant demographics (sample size and age); interventions for remote digital health management; intervention duration; involvement of health care professionals; measures to ensure participant compliance; and outcomes. The corresponding authors were contacted to clarify or obtain any information that was unclear or missing.

FXJ and WJN performed the quality assessment independently. Any disagreements were resolved through discussion between the 2 researchers, and a third investigator independently reviewed the final decisions. The quality of the RCTs was evaluated by 2 researchers using the Cochrane Risk of Bias Assessment Tool (ROB1) for assessing risk of bias. This tool addresses 6 domains of bias, including selection bias, performance bias, detection bias, attrition bias, reporting bias, and other biases. The risk of bias was classified as high, low, or unclear, with reasons provided for each classification.

When all the indicators are measured using the same standard, we aggregate the continuous data through the mean difference (MD), such as NS; while when different measurement standards are used, we use the standardized mean difference (SMD) for aggregation, such as QoL.

Considering the clinical heterogeneity in the population characteristics, specific contents of the intervention measures, and outcome measurement tools included in the study, we predetermined to adopt the random effects model as the main analytical framework [22]. This choice is based on conceptual assumptions rather than solely relying on the results of statistical heterogeneity tests. To improve the accuracy of effect size estimation in small samples or in the presence of heterogeneity, we used the Hartung-Knapp-Sidik-Jonkman (HKSJ) method to combine the random effects model and calculate the 95% CIs of the pooled effect size [22]. Effect sizes were interpreted using Cohen d criteria; values of 0.2, 0.5, and 0.8 represent small, medium, and large effects, respectively [23]. This systematic review uses restricted maximum likelihood estimation to estimate between-study variance, thereby more accurately reflecting the impact of true heterogeneity on effect estimates. To assess the heterogeneity among the studies more comprehensively, we used multiple indicators for quantification. For meta-analyses involving at least 3 studies, we reported the 95% CIs and 95% prediction intervals (PIs) to convey the range of effect sizes that new original studies might exhibit in different clinical scenarios in the future. For analyses with fewer than 3 studies, since the PIs might be too wide and lack information, they were not reported [24]. Meanwhile, we reported the I² statistic for comparison with previous studies but were cautious in interpretation, mainly by integrating τ² for a comprehensive judgment [22]. The standards recommended by the Cochrane Handbook are as follows: (0%‐40%) potentially insignificant heterogeneity; (30%‐60%) possibly moderate heterogeneity; (50%‐90%) possibly significant heterogeneity; and (75%‐100%) highly heterogeneous. It should be noted that these ranges are indicative only. The actual interpretation requires a comprehensive assessment considering the magnitude and direction of the effect size, the P value from the heterogeneity test (typically with P<.10 as the significance threshold), and the degree of overlap in the CIs presented in the forest plot [25]. To explore potential sources of heterogeneity, we conducted subgroup analyses based on intervention duration, intervention methods, and scale differences. To ensure the robustness of the results, a sensitivity analysis was conducted in this study (Multimedia Appendix 2). All statistical analyses were performed using R (version 4.5.2; R Foundation for Statistical Computing).

By drawing funnel plots and combining with Egger regression intercept test, the influence of small sample effects on the results was evaluated. It should be emphasized that such methods mainly detect small sample effects rather than directly testing publication bias [26,27]. Small sample effects may be caused by various factors, including publication bias, selective reporting, true heterogeneity among studies, and methodological differences, etc. According to methodological standards, funnel plots only have reliable interpretative power when at least 10 studies are included; otherwise, their interpretability is relatively low [26,28]. We applied Egger regression intercept test to all outcome indicators. It should be noted that when the number of included studies is small, statistical power will significantly decrease, which may lead to the generation of false negative results [27]. The outcome indicators finally published in this review are completely consistent with the prespecified plan.

To appraise the certainty of evidence, we applied GRADE (Grading of Recommendations Assessment, Development, and Evaluation) guidance (GRADE Working Group) to the main outcomes (NS and QoL). RL and WKZ reviewers independently rated evidence across risk of bias, inconsistency, indirectness, imprecision, and publication bias, and resolved disagreements by consensus.

The PRISMA flow diagram (Figure 1) illustrates the screening process and the criteria used for excluding papers. The initial search yielded 1608 publications. After deduplication using EndNote (Clarivate) software, 408 duplicate publications were excluded, leaving 1200 publications for title and abstract screening. Based on inclusion and exclusion criteria, 1171 publications that did not meet requirements were excluded, leaving 29 publications for full-text review. During the full-text review process, 8 studies were excluded due to incomplete data (n=4), irrelevant research subjects (n=2), or only one research group (n=2). Ultimately, 11 RCTs were included in the quantitative synthesis analysis [29-39].

The 11 included studies originated from 7 countries, including Korea (n=3) [29,30,38], Thailand (n=1) [31], Spain (n=1) [32], the United States (n=1) [33], China (n=3) [34,36,37], the Netherlands (n=1) [35], and Vietnam (n=1) [39]. Publication dates spanned 2018-2026, with a total sample size of 1210 participants: 598 in the intervention group and 612 in the control group [29-39]. Participants' ages ranged from 37.7 to 69.0 years. Digital health interventions encompassed diverse formats, including lifestyle interventions via mHealth apps [29,32-35,37-39], smartphone-based exercise programs [30], and nutrition education programs based on mobile apps, online platforms, and social media [31,36], with intervention durations ranging from 2 to 48 weeks. Of the 11 studies, 8 [29-31,34-37,39] explicitly mentioned health care professionals (eg, coaches, nutritionists, nurses, and doctors) participating in the intervention process. Eight studies used specific strategies to enhance intervention adherence. Among them, 3 studies [29,33,35] used automated text messages or app push notifications sent by digital platforms according to preset schedules, with the content mostly consisting of standardized meal check-in reminders or general health tips. Four studies [30,34,36,37] involved follow-up calls or WeChat (Tencent Holdings Limited) messages from professional health care providers, along with face-to-face evaluations. These interactions typically featured personalized consultations, and researchers would provide targeted advice based on the specific data of the participants. One study [38] included both text messages sent by the digital platform and phone follow-ups by health care providers to enhance patient compliance. Regarding outcome measures, 10 studies [29,30,32-39] included QoL assessments, and 2 studies [31,36] conducted NS assessments. Detailed baseline information for included studies is presented in Table S1 in Multimedia Appendix 3.

This study strictly adhered to the PRISMA guidelines, with 11 RCTs ultimately included. The Cochrane RoB 1.0 was adopted to evaluate the quality of the trials across 6 dimensions. Figures 2 and 3 illustrate the assessment of bias risk in RCTs. Among the 11 included studies, most had high quality in terms of attrition bias, reporting bias, and other biases, all rated as low risk [29,30,32-39]. However, there were many methodological issues in terms of selection bias and implementation bias. Specifically, in the generation of random sequences, 1 (9%) study [31] was rated as high risk due to the use of non-random methods such as order of admission, 5 (45%) studies [32,34,36-38] were rated as unclear because they only mentioned randomization but did not describe the specific generation method; in the aspect of allocation concealment, 9 (82%) studies [30-34,36-39] were rated as unclear because they did not report the implementation details of allocation concealment (eg, sealed envelopes or central randomization); in the aspect of participant and personnel blinding, 1 (9%) study [30] was rated as high risk due to the lack of implementation of blinding or the blinding being compromised, and 8 (73%) studies [29,31-34,36,37,39] were classified as unclear due to the incomplete description of the blinding situation in the reports; in the aspect of outcome assessment blinding, 5 (45%) studies [30,31,33,36,37] were rated as unclear because the reporting did not clearly describe the situation of outcome assessors knowing the group allocation. The dropout rate in this study was low and the groups were well balanced, indicating that the risk of dropout bias was within an acceptable range.

In terms of management intervention measures, 8 trials [29-31,34-37,39] involved the participation of medical professionals in using digital health tools, while the remaining 3 [32,33,38] did not provide specific information (Table S1 in Multimedia Appendix 3). The main roles of medical professionals include developing personalized intervention plans [29,37], providing nutritional guidance and exercise training [29,31], guiding device use [30,37], reviewing health data [29,35], and conducting subsequent question answering and plan adjustments [34,36]. In all studies, a total of 11 different digital health tools were used, covering 6 mHealth apps, 2 web programs, 2 smart devices, and 1 tool that combines a web page with WeChat. These tools were mainly applied in 2 key areas of CRC management, including self-management [30,31,33,36-39] and rehabilitation support [29,32,34,35]. Self-management platforms typically have functions such as health knowledge dissemination, data recording, goal setting, and question answering; rehabilitation support platforms focus on tracking postoperative recovery data, monitoring physical activities, and enabling real-time viewing and intervention by the medical team on the rehabilitation status. Some tools require the use of smart devices (eg, wristbands) and can display health information in real time through the connection to the device screen. An overview of the characteristics of digital health tools is shown in Table S2 (Multimedia Appendix 3).

To assess the impact of small studies, we combined visual inspection of the funnel plot with Egger regression intercept test in accordance with methodological guidelines [40]. We constructed a funnel plot for QoL (Figure 9) to visually explore potential asymmetry and did not detect a significant influence of small studies (z=1.29; P=.20). However, because only 10 studies were included, the symmetry of the funnel plot or the nonsignificant result of the Egger test cannot rule out the presence of publication bias. Moreover, if asymmetry is observed in the funnel plot, it may be due to heterogeneity or methodological flaws rather than solely publication bias [40]. Therefore, the current results do not rule out the possibility of publication bias or selective reporting bias, as the number of included studies is small, and these biases may not have been detected yet.

Based on the GRADE evidence grading method, the quality of evidence regarding the effect of digital health interventions on improving NS and QoL in patients with CRC was evaluated. For the outcome of NS, 2 RCTs were included [31,36]. Due to serious risk of bias (inadequate reporting of randomization and blinding, reliance on subjective scales for outcome measurement), significant inconsistency (large discrepancies in effect estimates across studies), and serious imprecision (the 95% CI crossing the line of no effect and a small sample size), the overall quality of evidence was rated as low, indicating that the true effect is highly uncertain and subsequent research is very likely to change the current conclusion. For the outcome of QoL, 10 RCTs were included [29,30,32-39]. The risk of bias was acceptable, and no downgrading was applied for indirectness or imprecision. However, some heterogeneity existed across studies (variations in effect sizes due to differences in intervention methods, duration, and measurement tools), leading to a one-level downgrade for inconsistency. The final evidence quality was rated as moderate, with the pooled effect showing that digital health interventions can moderately improve patients’ QoL. This suggests that the true effect is likely to be close to the estimate, though some possibility remains that future studies may modify the finding. Results are summarized in Table 1.

This study conducted a meta-analysis on the effects of digital health interventions on the NS and QoL of patients with CRC. Regarding NS, based on 2 studies using albumin as an indicator, the analysis showed that digital health interventions did not significantly improve the NS of patients with CRC. The pooled effect size was not statistically significant, and there was high heterogeneity between the studies, so the results should be interpreted with caution. Regarding QoL, a comprehensive analysis of 10 studies indicated that digital health interventions significantly improved the QoL of patients with CRC overall, with a moderate effect size. However, there was high heterogeneity among the studies, and the prediction interval was wide, suggesting that the intervention effect may fluctuate considerably across different populations. To explore the sources of heterogeneity, further subgroup analyses were conducted; the intervention duration subgroup showed that short-term intervention had no significant effect, while medium-term and long-term interventions had significant effects, but the differences among the subgroups were not statistically significant. The intervention method subgroup revealed that supervised interventions had a stable and significant effect, unsupervised interventions had a minimal effect, and combined interventions had an unstable effect. The difference between these subgroups was significant, indicating that the type of intervention is an important effect modifier. The QoL scale subgroup showed that different measurement tools had some influence on the effect size, but the difference between subgroups was not significant. In conclusion, digital health interventions have positive potential for improving patients’ QoL, especially in supervised intervention models where the effect is more pronounced. However, their effect on improving NS remains unclear, and more high-quality research is needed for further verification in the future.

The quality of the included literature was generally high, with a low risk of bias. However, most studies did not mention the use of assessor blinding in outcome measurement [30,31,33,36,37]. Future research should aim to improve the design related to blinding. In the GRADE assessment, the evidence level for NS was rated as low [31,36], and the evidence level for QoL was rated as moderate [29,30,32-39]. Even after conducting sensitivity analyses, heterogeneity persisted across studies. The risk of bias for all outcome measures was downgraded due to issues with allocation concealment and inadequate implementation of blinding methods [29,30,36,37,39]. Consequently, researchers should prioritize standardizing experimental designs and meticulously executing these processes in future studies.

The innovation of this study mainly lies in the following aspects: first, it focuses on the application of digital health intervention in the specific group of patients with CRC and examines its dual effects on NS and QoL, filling the gap in existing research where insufficient attention was paid to nutritional outcomes. Second, through detailed subgroup analysis, the study first clearly indicates that the intervention method (especially the guidance-based intervention) is the key regulatory variable leading to differences in effects, and reveals that the guidance-based intervention can stably improve the QoL in the case of zero heterogeneity, providing precise directional guidance for the design of subsequent intervention plans.

This study also has certain limitations. First, although 11 RCTs were included, the overall sample size is relatively limited, and there are significant differences in intervention duration and outcome measurement tools among the studies, resulting in high heterogeneity in the combined analysis, which to some extent weakens the robustness of the conclusion. Second, in the evaluation of NS, the study only uses the albumin level as the core indicator, which is relatively single and fails to comprehensively reflect the complex NS of patients (weight changes, muscle mass, dietary intake, etc), and the high heterogeneity of this outcome suggests that potential confounding factors may not have been adequately controlled. Third, the search strategy of this study is limited to published literature. It does not systematically search for gray literature and unpublished data from clinical trial registration platforms. This limitation may lead to potential publication bias, that is, some negative or unpublished results may be missed. Future research should adopt larger-scale and longer-term investigation methods and use standardized assessment standards to implement these standards in different age groups to further verify these findings.

This systematic review indicates that digital health interventions show certain potential in improving the QoL of patients with CRC, with a combined effect size at a medium level. However, the interpretation of this result should be done with caution. First, the quality of evidence is limited by the high heterogeneity of the included studies, with a wide prediction interval, suggesting that the effect may not be significant or even have a negative effect in specific populations or intervention contexts. The stability and universality of the effect still need to be verified. Second, subgroup analysis reveals that the improvement in intervention effect may be related to specific design features, especially the guidance-based interactive mode and an intervention period of at least one month, indicating that not all digital interventions can produce equivalent benefits. The self-monitoring mode relying solely on technical means has limited effect. Finally, there is still a lack of sufficient evidence for the effectiveness of digital interventions in improving objective nutritional indicators (eg, albumin levels) of patients. Therefore, future research should not stop at verifying the overall effectiveness but should further focus on exploring the response differences of patients with different characteristics (eg, age, digital literacy, and disease stage) and develop evidence-based, more adaptable stratified intervention strategies in order to achieve truly precise health management in the future.