Within the 27 included studies, we identified 26 unique ePROM-integrated/self-monitoring intervention programs (either developed, planned, or simulated). The key characteristics of these interventions are (1) the country/setting, (2) the intended users (by one or multiple cancer types), (3) the ePROM tools or frameworks embedded in the intervention design, and (4) the symptoms monitored by a particular intervention. Table 1 displays a snapshot of the most commonly identified characteristics of each intervention, listed in order of the number and percentage of interventions with a particular characteristic. The following properties are non–mutually exclusive, as a particular ePROM program may have been developed for multiple countries and cancer populations, and/or use a combination of different ePROM tools, etc. A representative median ePROM intervention would be developed in a single country and use 1-2 validated ePROM tools to monitor around 6-7 symptoms. However, the symptom lists extracted from the studies may be incomplete and therefore undercounted. Most interventions catered to a patient population characterized either by a single cancer site (eg, breast) or by an umbrella term (eg, all cancers or solid cancers), but some interventions catered to up to 4 distinct cancer groups.

Table 1 shows a variety of countries involved in the research and development of ePROM interventions, mostly comprising European Union members, the United Kingdom, the United States, and Australia. The types of cancer represented most within the identified interventions are breast and colorectal. Except for the 7 studies that developed novel ePROM frameworks, most interventions incorporated established published tools, such as Patient-Reported Outcomes Version of the Common Terminology Criteria for Adverse Events (PRO-CTCAE), European Organization for Research and Treatment of Cancer (EORTC) Quality of Life Questionnaire (QLQ)-C30, and the Distress Thermometer; 1 study did not report this information [40-42]. Most interventions were designed to monitor an extensive list of symptoms, most commonly including fatigue, pain, and nausea. The symptom lists included in the ePROM intervention were not identified in 7 studies.

We grouped the validated ePROM instruments embedded in the included studies by aim and functionality into the following 7 categories.

First, many of the instruments were symptom screening tools that measured (1) presence and severity, and (2) the distress or interference caused by each symptom. This category included the Advanced Symptom Management System (ASyMS), the Brief Pain Inventory (BPI), the Chemotherapy Symptom Assessment Scale (C-SAS), the Chemotherapy Toxicity Self-Assessment Questionnaire (CTAQ), the Memorial Symptom Assessment Scale (MSAS), the PRO-CTCAE, and the Patient-Reported Outcomes Measurement Information System (PROMIS) [40,44,47-51].

The second category differed subtly from the first by comprising symptom screening tools that measure only symptom presence and/or severity, but not the distress level or the impact of a symptom on daily life or experience. This category included the Checklist Individual Strength, the EORTC QLQs, the Edmonton Symptom Assessment Scale (ESAS), Hospital Anxiety and Depression Scale (HADS), the Personal Health Questionnaire (PHQ) Depression Scales (PHQ-8 and PHQ-9), STAR, and the World Health Organization Well-Being Scale-5 (WHO-5) [41,52-59].

Four categories consisted of measures that focus on broader health-related aspects than symptoms. First, some tools incorporated generic measures of overall health, HRQoL, or distress level and/or problem checklists. These included the Distress Thermometer and Short Form Health Survey-36 (SF-36) [42,60]. Second, some of the questionnaires measured satisfaction with treatment and support, unmet needs, fear of cancer recurrence, and/or other patient-reported experience measures. These measures included HRQoL and overlapped with other outcomes. This category included the Breast Questionnaire, the Cancer Worry Scale-6, the Holistic Needs Assessment (Health MOT; an “MOT” refers to a British vehicle safety test named after the now-defunct Ministry of Transport), and the Supportive Care Needs Survey-ST9 [45,61-63]. Third, some questionnaires measured patient-reported behaviors using nutrition questionnaires, smoking and alcohol abstinence scales, and the Short Questionnaire to Assess Health-Enhancing Physical Activity [64-66]. Finally, some questionnaires measured patients’ knowledge, ability, and understanding, such as the Digital Health Literacy Instrument and the European Health Literacy Survey [67,68].

It is important to note that the listed tools were not entirely separate from each other, as some of the instruments were based on a mix of previously developed frameworks; for example, ASyMS integrated the Common Terminology Criteria for Adverse Events grading scale with the C-SAS instrument [48].

Table 2 presents an abridged version of the SLR study summary table, the corresponding detailed version of which is presented in Table S1 in Multimedia Appendix 1. The full supplementary table contains detailed summaries of the included studies (n=27), listed alphabetically either by the study or project name or by the authors’ names if the study name was not available. As such, for the studies with multiple publications included in this review, data were extracted collectively by cross-referencing the grouped publications, rather than extracting each publication separately. This helped the extraction of complementary information to gain insight into the link between detailed ePROM characteristics and health economic outcomes. For example, the Sentinel HEA report did not include the detailed information on the ePROM intervention design and procedures reported in the main RCT paper [34,35].

This review identified several distinct design frameworks for ePROM-integrated interventions from studies with health economic components. However, there was a considerable variation in the level of technical detail with which the evaluated digital system was described. In most cases, the intervention was designed according to the following pattern: (1) the patient remotely reported symptoms and completed ePROMs; then (2) the ePROM information triggered alerts or was processed into an action recommendation; after which (3) the ePROM information, the alert, and/or the action recommendation were sent to a designated agent responsible for deciding on further actions. The systems can be categorized by the actor responsible for making decisions following ePROM processing (“active” vs “reactive”), the extent to which decisions are automated, the type of service offered to patients following a decision, and any additional functionality fostering patient-HCP communication (eg, within-app messaging or consultation scheduling). In addition to the active and reactive approaches, some interventions do not use alerts and instead treat ePROM information as supplementary to support HCPs in making treatment decisions or structuring consultations.

In the active approach to decision-making and alert handling, ePROMs and/or ePROM-triggered alerts are sent to HCPs for actioning, whereas the reactive approach may entail sending ePROM-tailored self-help information and/or guidance on recommended actions to patients [82,84]. Examples of strictly or predominantly active-approach interventions are AmbuFlex, eHealth@Hospital-2-Home, ICECARE-P, MyPal, PROMISE, PROMPT-Care, Sentinel, and STAR [5,35,38,71,73,74,77,81]. Reactive-approach interventions include Affordable Better, CAG, CAPRI, PRO-NET, and Oncokompas [36,65,70,78,82]. Some interventions blend both methods into a hybrid approach, usually to distinguish between alert severity. In these cases, low-priority alerts are handled reactively by sending patient self-help information tailored to their symptom reports. In contrast, high-priority alerts are actively sent to HCPs with urgent action recommendations. Such “hybrid” approaches are present in the eRAPID, eSMART, LYSA, PHONEME/Interaktor, PRISMS, SIS.NET, and YES interventions [33,50,76,79,80,83,87]. Interestingly, the SYMPRO-Lung protocol outlines a plan to explicitly compare the active and reactive approaches in separate trial subgroups. For interventions that provide self-help information and guidance, it is important to distinguish those that are predominantly active, but allow patients to access information should they choose to do so independently of ePROM results (eg, eHealth@Hospital-2-Home), from those that are “proactively” reactive and send patients information linked to or triggered by their ePROMs (eg, CAG and PRO-NET) [65,73,82]. The intervention presented by Ferre et al [69], as well as VBHC, did not use alert handling and instead treated ePROM data as supplementary information for HCPs [69,86]. In the publications from Kaiku Health and the Programme, it is not explicitly clear how/if alerts are handled [75,85].

Examples of additional functionality embedded in the ePROM interventions, which contribute to patient-HCP communication at the cost of increasing intervention complexity, are in-app/portal appointment scheduling (eg, APCOT), in-app/portal patient-HCP videocall and messaging (eg, Affordable Better), and automated ePROM data summary dashboards displayed to HCPs and/or patients (eg, CAPRI, Ferre et al [69]) [36,70,72].

We identified 10 studies reporting economic results, grouped into full (cost-effectiveness), partial (cost-consequence), or HRU-only analyses. We visualized each study by (1) outcome, where “Positive” indicates ePROM-monitoring is associated with low incremental cost-effectiveness ratios (ICERs), cost savings, higher certainty, etc; (2) alert handling protocol, where “Responders” refers to the designated ePROM-informed decision maker; and (3) study sample size, shown in Figure 3.

Only 5 studies reported cost-effectiveness results, including ICERs (5/5 studies) and/or cost-effectiveness acceptability curves (3/5 studies), as shown in Table 3. From these, 2 studies (Sentinel and STAR) reported strong evidence for high ePROM cost-effectiveness relative to usual care, while 3 studies (eRAPID, PHONEME/Interaktor, Oncokompas) reported positive but more uncertain results [6,34,39,78,79].

A total of 5 other studies (CAPRI, Affordable Better, the Programme, PROMPT-Care, and SIS.NET) reported reductions in costs or resource use within partial economic analyses alongside positive clinical or implementation results (Table 3). From these, 4/5 studies could have potentially yielded ePROM strategy dominance and/or high probabilities of cost-effectiveness following a full HEA [5,37,70,85]. On the other hand, 1 study (SIS.NET) did not support the hypothesized reduction in HRU for ePROMs; it thus might have yielded a result with a high and uncertain ICER and/or dominant standard care strategy following a full HEA [83].

We provided detailed study-by-study descriptive summaries of cost-effectiveness results accompanying Table 3 in the Multimedia Appendix 1. Table 3 also presents the alert-handling approach (see the “Intervention Design” section) and cancer type in each study.

In summary, evidence of QALY gains or health utility improvements from ePROM interventions is provided in Sentinel, STAR, and PHONEME/Interaktor. STAR, CAPRI, and PROMPT-Care report evidence that remote monitoring reduces emergency visits, hospitalizations, and/or hospitalization duration. The Sentinel and CAPRI trials demonstrated that the increased independence of ePROM-tracked patients shifts their health-related travel patterns from expensive (eg, ambulances and taxis) to cheaper (eg, walking and driving) modes. Conversely, significant differences across other HRU categories (eg, outpatient appointments and tests/scans) were not detected in the studies included in this review.

Twelve studies included qualitative research components (Table S2 in Multimedia Appendix 1 for the detailed extraction). Of these, most (7/12) conducted their qualitative research alongside or at the end of a pilot study or trial, while 3/12 did so during the design/planning stage, and 2/12 included multiple qualitative components throughout various study phases. A total of 11/12 studies conducted interviews, of which 7 used a semistructured framework, 1 used a think-aloud approach, while 3 did not specify the interviewing style. Overall, 3/12 studies organized focus groups and workshops, while 2/12 collected qualitative survey or feedback form data. Only 6/12 studies explicitly specified the type of qualitative analysis, which comprised thematic (n=5) and content (n=1) analyses.

For each of the 12 studies with qualitative components, we extracted the aims, participant numbers, and types (eg, patients, HCPs, or both), conclusions (if applicable), and a list of themes analyzed (Table S2 in Multimedia Appendix 1). In most instances, authors explicitly listed the qualitative themes present in their research. Still, we supplemented this in our extraction based on our reading of the studies or when thematic descriptions were limited. Only 1 study (VBHC) did not discuss or report its thematic approach [86].

We counted and categorized instances of each theme type in the 12 studies. The most common theme was “user experience,” included in 6 studies. This was followed by “acceptability,” “impact on care,” “self-management,” “adherence/adoption,” “implementation,” each of which featured in 3 studies. Then, “responses to alerts” and “usability” each appeared twice, while 7 studies included “other” themes. The “other” category consisted mostly of design- or implementation-related themes; for example, discussions on design recommendations or intervention fidelity (ie, “Was the intervention implemented as intended?”).

We evaluated the relevance of the identified themes to this review based on (1) how directly related they are to cost-effectiveness–related questions about ePROM integration and (2) how useful the insights from their analyses are in formulating health economic modeling assumptions and/or sensitivity/scenario analyses. Following our extraction of quantitative data, we formulated the following example questions that we hoped our qualitative extraction would address: How does ePROM integration impact time use and workloads? How/why does the design of the intervention (eg, active vs reactive alert handling) impact resource use?

However, the most common themes present in the included qualitative analyses were of limited health economic relevance and were instead related to app/platform users (eg, “user experience,” “acceptability,” and “usability”). These, in turn, involve discussions with users (eg, patients or HCPs) of an ePROM app or platform about its usefulness in health care–related communication, personalization, feedback features, and visual/interface design. In most of the identified papers discussing this theme, the authors report acceptance and enthusiasm from users related to these ePROM attributes, but do not link these findings to any health economic insights.

Two studies (VBHC and PHONEME/Interaktor) used qualitative methods to guide the quantitative outcome and HRU variables and questionnaire design [79,86].

Only 2 studies conducted qualitative research specifically related to health economics. In eRAPID, discussing the impact of the intervention on consultation time yielded mixed results, with HCP opinions divided between ePROMs being “possibly time saving in some instances” and ePROMs “adding time to the consultation.” However, the authors did not focus on this aspect or provide further detail in the report [33]. The PHONEME/Interaktor study qualitative components comprised (1) HRU variable identification (to guide and design the cost-effectiveness analysis) and (2) interviewing nurses about the impact of alert handling on workloads [79]. The latter component is by far the most relevant type of qualitative research to the aims of this SLR. Unfortunately, it is part of a separate, as-yet-unpublished manuscript, and therefore could not have been extracted for this review.

Within the 10 studies reporting quantitative health-economic results, a high degree of heterogeneity in setting and intervention design obscured links between these characteristics and health-economic measures. The 2 studies that presented the most promising and complete evidence for ePROMs’ cost-effectiveness (Sentinel and STAR, see Table 3) were both active interventions, albeit set in different contexts (France and United States/Canada, respectively). Nevertheless, reactive interventions such as Affordable Better in the Netherlands and CAPRI in France demonstrated cost savings. As we discussed above, the evidence for other interventions (eg, hybrid-approach eRAPID) ranges from high uncertainty to a demonstration of ePROMs’ ineffectiveness.

We hoped to identify qualitative analyses explaining how and why active and reactive ePROM alert-handling approaches affect health outcomes and resource use. However, none were included in this review’s extraction. Therefore, we currently have limited insight into the effect of ePROM design on cost-effectiveness, but we have identified a study planning to address this gap in the future (SYMPRO-Lung) [84].

Key parameters of interest for extraction in this review are HRU rates among patients using ePROM interventions, as well as corresponding rates and incidence rate ratios (IRRs) for comparison with standard (in-person) care. Ideally, these would be extracted from appropriately adjusted regression models from RCTs, but most studies did not report their results in this way. Papers predominantly reported raw, unadjusted counts of observed events per study group during follow-up, making comparisons across studies difficult.

However, using separately reported study follow-up information (eg, from the CONSORT [Consolidated Standards of Reporting Trials] diagrams), we calculated person-time and converted raw HRU counts into monthly rates for cross-compatibility and easy use in future economic modeling through the actuarial life-table approach given by , where: [89].

For example, the CONSORT diagram presented in Absolom et al [4] shows the eRAPID arm having 256, 239, 222, and 214 patients on study at baseline, 6 weeks, 12 weeks, and 18 weeks, respectively, yielding 4176 person-weeks (≈960 person-months) when inserted into the corresponding equation. An excerpt of HRU and ePROM alert counts converted into monthly rates (events/person-month) is presented in Table 4; the full extraction is available in Table S3 in Multimedia Appendix 1.

We calculated IRRs in Table 4 from available person-time and event rates or frequencies reported in the source paper. IRRs can be used to approximate hazard ratios if (1) the average follow-up is comparable between groups and (2) the relative hazards are constant [91]. A limitation of this approach in this context stems from a difficulty in verifying these assumptions (especially the latter) and the inability to account for competing risks using the cause-specific Cox proportional hazards model on individual-level data (as would be the case in a preplanned analysis) [92]. For example, the Sentinel trial followed patients in both arms for 587 days (from June 1, 2014, to January 9, 2016), broadly satisfying the first assumption despite an observed survival improvement in the ePROM arm resulting in slightly more person-months experienced than in the control arm (631.4 vs 601.4 person-months) [35]. However, HRU events reported by the study were discussed through the prism of differences between arms rather than over time, so this information is missing; it is reasonable to suspect that the amount of HRU events experienced by patients with lung cancer after or during treatment may change over time, thus violating the constant hazards assumption [34]. The method nevertheless yields approximations of useful modeling and sensitivity analysis parameters without access to sensitive patient-level trial data. We calculated SEs and CIs for IRRs using the following formulae (with E as the number of events observed in the ePROM=1 and control=0 groups) [93].

The IRRs derived from the HRU monthly event rates reported in the CAPRI trial are subject to additional uncertainty, as the authors did not report total event rates and were unable to provide them upon request [37]. However, the paper does state that the mean follow-up time used to calculate these rates was 4.58 months [37]. For crude IRR SE and CI estimation, we approximated the event count (E) values by multiplying the trial arm populations by the rate and the mean follow-up: N × λ × 4.56, a method endorsed by the authors (after contacting them). The estimated IRR CI widths broadly align with the reported P values for rate differences in the paper and support the authors’ conclusion of a significant reduction in hospitalizations with ePROM remote monitoring. However, IRRs are more useful than additive rate differences for economic model parameterization and sensitivity analysis.

The event rates IRRs from the SIS.NET study were not calculated on a per-person-time basis, as this was not reported in the paper’s CONSORT diagram [83]. Instead, the authors present mean per-patient 18-month values. The approximated monthly figures in Table 4 were derived by dividing the reported values by 18, so IRR SEs and confidence bounds could not be calculated.

The standard care arm and corresponding IRRs from the Affordable Better extraction are absent because the control arm was not empirical; the authors created a hypothetical counterfactual control based on the ePROM arm and assumptions about differences between hypothetical study arms in in-clinic and remote appointments [70].

The STAR monthly death rates were calculated from the reported adjusted 1-year survival results (p) using the probability-to-rate conversion formula, where t=12 months (notes: constant hazard assumed, inverse CI bounds were swapped) [22,38]. The study presented statistically significant survival gains in the ePROM arm relative to control; the effect was greater in the computer-inexperienced subgroup [38].

Populating health modeling parameters requires health state utility values, rather than cumulative QALYs. EQ-5D–derived health utilities by trial arm and time point are reported in the eRAPID and STAR papers [33,38]. However, the eRAPID report only presented descriptive statistics (means and SDs) for EQ-5D scores, while the STAR paper reported CIs and adjusted between-arm differences and P values; STAR utilities were converted from their original 0-100 scale to a 0-1 scale in this extraction [33,38]. We presented extracted EQ-5D health utility data in Table 5.

In the ePROM results, similar health utility patterns between arms are consistent with a lack of statistically significant QALY differences detected by the multivariate regression analysis. Conversely, STAR’s point drop differences can be interpreted as health utility gains due to effective ePROM monitoring; the increased (albeit less statistically powerful) difference in the computer-inexperienced patients is consistent with the subgroup’s overall more favorable results.

Potentially important parameters for economic modeling of ePROM interventions include decision-tree or transition probabilities related to alert handling; for example, what is the probability that a patient is referred to the hospital after a medium/moderate alert? These are separate (but related) to the frequency rates of alerts presented in Table 4. Such probabilities can be derived from the results of the AmbuFlex and STAR reports.

AmbuFlex elicited 374 responses from patients with prostate cancer between September 2014 and December 2015, of which 38 (11%) were classed as “green” by the algorithm (ie, no further contact required), 128 (37%) were “yellow” (triage), and 181 (52%) were “red” (clinic referral); from the presented results, it is possible to infer that 52 (41%) of the “yellow” reports did not require further action, while 76 (59%) led to clinic referral [71].

The total number of individual symptom reports (n=84,212, of which 1431 were graded ≥3) is presented in the STAR intervention, but not the total number of email alerts to nurses (when symptoms worsened by ≥2 points or reached an absolute grade ≥3). However, simple percentages of nurses’ responses were reported: telephone counseling (77%), supportive medication initiation/change (12%), hospital/emergency referral (8%), chemotherapy dose modification (2%), and imaging/test orders (2%) [38]. Exact absolute values corresponding to these percentages are missing from the report and cannot be derived due to rounding; furthermore, these percentage sums slightly exceed 100% (also due to rounding) and would need to be adjusted for use in model parameterization.

In PROMPT-Care, authors present detailed breakdowns of the 877 alerts, with 44% (383/877) reviewed by HCPs, resulting in 496 actions. The action outcomes included no further follow-up (83/496, 17%), no response (111/496, 22%), or an in-clinic or remote follow-up appointment (302/496, 61%); follow-up actions comprised HCP-patient discussions (129/302, 43%), information provision (98/302, 32%), or onward referral (75/302, 25%) [5].

We identified 2 main cost categories of relevance to ePROM monitoring: (1) upfront setup, including IT, equipment, and web/app development, and (2) maintenance costs, comprising HCP wages, IT administration, etc. However, cost reporting was inconsistent across studies. We identified upfront cost estimates ranging from GBP 63-GBP 102/patient (GBP 1=US $1.3635 as of August 24, 2028) to GBP 141K (total), and running costs ranging from GBP 37/patient-year to GBP 217/patient-month (reported in 2024 GBP) [34,39,85,86].

We have included summaries of ePROM monitoring costs and details of currency conversion and inflation adjustments in the Multimedia Appendix 1.

For parametrizing a health economic model that incorporates ePROM-elicited information (eg, symptom severity, see Figure 1), information about the overall difference in rates of HRU between treatment strategies (as most concluded studies report) may be insufficient; it is also necessary to know the relationship between ePROM data and HRU (eg, given that a patient self-reported a high pain score, can they expect to be hospitalized sooner or more frequently?). However, this link is scarcely reported in the included studies.

The SIS.NET study reported the most direct link between ePROMs and HRU in the form of a direct positive correlation between symptom reporting rates and non–breast cancer–related appointments; however, only a P value (P=.0004) is provided, so no formula is available for use in a model [83].

A detailed, yet indirect, link between ePROMs and HRU rates is reported in the PHONEME/Interaktor study’s multivariate regression analyses [79]. The link is “indirect” because the patient-reported independent variable included is baseline EQ-5D health utility rather than disease-specific ePROM data (eg, symptom reports); with the right mapping algorithm and assumptions, a modeler may be able to use these formulae to predict HRU rates based on symptom scores. However, the regression tables only reported coefficient estimates without intercept values, rendering the formulae nonreconstructible; however, the authors provided the full results tables upon contact, which are provided in this paper’s Multimedia Appendix 1. Despite the absence of a covariance matrix, we performed a secondary analysis alongside this review based on the PHONEME/Interaktor formulae to demonstrate their use in a hypothetical model’s probabilistic sensitivity analysis. Example results of various HRU rate predictions for a patient with breast cancer aged 60 years after 2 neoadjuvant chemotherapy cycles with a Charlson Comorbidity Score of 1 are presented in Figure 4. The full formulae and the code used to perform this analysis are provided in the Multimedia Appendix 1.

This review identified 27 unique complex digital interventions that embed patient-reported outcomes into clinical practice and decision-making. These elicit patient-reported information from either a variety of validated ePROM tools or novel frameworks. Most (but not all) systems include alert handling and/or automated decision support based on alert classification, using either an active or reactive approach to choosing an agent responsible for taking recommended actions.

We did not identify any form of guidance, methodology, or best practice specific to modeling oncological ePROM interventions.

There is mixed evidence on the cost-effectiveness of ePROM interventions, further complicated by the heterogeneity of intervention designs. However, out of the 5 studies that conducted cost-effectiveness analyses, 3 showed neutral or uncertain results, while 2 showed strong evidence in favor of integrating ePROMs into clinical practice.

Only 1 health economic model was identified in this review. However, it followed a simple 2-state Markov chain structure and is of limited use to potential future modelers aiming to adapt it or run more complex simulations or scenario analyses in different contexts. While such a model benefits from parsimony, its parameters are elicited from simple differences between ePROM and usual care trial arms and therefore could not be easily used in a simulation of a different intervention with an alternative design framework or alert handling mechanism.

Qualitative evidence that could support future economic modeling assumptions was searched within the included summaries. However, this was seldom present, with the qualitative focus centering more on user experience and intervention design-related themes. Only 1 study (eRAPID) explicitly reported limited economic-specific findings, indicating possible time savings for participating HCPs thanks to automated ePROM data summaries.

The detailed parameter extraction we produced in this review, along with our recommendations for future research, could help modelers simulate the cost-effectiveness and test scenario analyses for a wide variety of ePROM interventions with heterogeneous design frameworks. For example, mapping relationships between self-reported symptoms and HRU patterns or generic HRQoL measures could enable the creation of early health economic models of upcoming ePROM systems that simulate the impact of different alert-handling and response mechanisms on long-term cost-effectiveness.

All of the economic evidence of ePROM monitoring we identified in this review comes from high-income countries where internet usage rates exceed 90% [94]. Most of these are European states with government-run or socialized health service providers.

With only 10 studies reporting heterogeneous economic results, it is difficult to infer the impact of other contextual aspects on PROM monitoring’s cost-effectiveness (eg, staffing or baseline service use). However, we did identify broadly favorable results for both active and reactive interventions with a variety of organizational setup structures (Figure 3), which designate the task of responding to ePROM-generated symptom reports to patients, nurses, or care teams, etc.

This review extracted and presented a heterogeneous set of HRU consumption rates, which could provide future modelers with the tools necessary to simulate differences in resource use and, therefore, costs between ePROM and standard care treatment strategies. Furthermore, this toolset was expanded by our secondary analysis of extracted results, in which we standardized all rates into comparable (monthly) units and produced IRRs. Event rates of patients with cancer by treatment strategy are well-evidenced and need not be a priority for future research relative to other gaps identified in this review.

The usefulness of the IRRs reported in Table 4 stems from enabling (1) the simulation of control populations from data collected in observational studies of ePROM interventions and (2) the construction of useful scenario analyses for economic models. For example, in a model scenario in which ePROM interventions increase patients’ independence in health-related travel (ie, more trips by public transport and fewer by ambulance), IRRs derived from the Sentinel study could be used. In contrast, CAPRI-derived IRRs could be used in scenarios where ePROM interventions lower hospitalization rates [34,37]. However, most of the IRRs in Table 4 were produced from this review’s secondary analysis of the included studies’ results and are therefore crude and unadjusted by any appropriate regression model.

We suggest that prospective modelers use our post hoc estimates with caution and care to match the context of interest. For example, IRR estimates from Sentinel may be most applicable to lung cancer populations in high-income settings with active alert-handling designs; extrapolating to other patient groups should be accompanied by transparent disclaimers of limitations and restricted to sensitivity/scenario rather than base case analyses.

To improve this evidence, future studies evaluating ePROMs and HRU consumption should include baseline-adjusted Cox proportional hazard or multivariate regression models using either ePROM study arm membership or granular self-reported symptom scores as predictors of HRU rates.

The extraction of utility values revealed a limited availability of data needed to build a health economic model. Table 5 includes health utility values for ePROM and standard treatment strategies, not for different health states per se (Figure 1). To effectively parametrize a model that tracked or simulated patients’ ePROM data (eg, symptom severity), a mapping algorithm from disease-specific ePROMs into generic measures such as EQ-5D-5L–derived health utility would be needed. A targeted PubMed search revealed validated algorithms for some ePROM instruments used by interventions included in this review (eg, EORTC QLQ-C30 and ESAS), but a lack of evidence for others (eg, PROMIS and novel instruments).

While we are not aware of any other ongoing methodological work in this area, we echo the recommendations provided by an SLR of mapping algorithms in rare diseases (including some cancers) [95]. Its authors (Meregaglia et al [95]) encouraged the development of algorithms with broader generalizability, improved quality, and higher robustness; they also cautioned future researchers against reproducing common errors their review identified (eg, overestimating health utility for severe health states, especially when using ordinary least squares regressions).

Developing mapping algorithms for a broader set of measures and symptom scores for ePROM profiles should be a priority for quality-of-life research in cancer care. It would also help future modelers effectively assign health utility values to symptom-severity–stratified health states. More useful yet would be a calculation of health utility by overall alert severity grade, which was not identified in any of the studies included or elsewhere.

Overall, this review failed to identify the kinds of qualitative research it searched for (ie, qualitative insights that could inform economic modeling assumptions or explain how or why ePROMs change HRU patterns). As a result, modelers may not have sufficient information to properly structure a decision-analytic design. We summarized the gaps and proposed research needed to address them in this section.

The conclusions of the completed qualitative research (in 6/12 studies) reported broadly positive feedback, enthusiasm, and potential related to the implementation and design aspects of the ePROM interventions. Other than (to a very limited extent) eRAPID, none of the studies presented qualitative results that could constitute health economic model structure guidance or inform modeling assumptions. To address this, a mixed methods approach could be taken in a future study, such as including qualitative interviews within a time-motion study, as shown by Singh et al [96].

Furthermore, some areas in our quantitative extraction would benefit from qualitative insight. Chiefly, Lizée et al [34] found that ePROM-monitored patients used cheaper, more independent forms of travel (eg, more walking and fewer ambulance trips), but did not explain the mechanism behind this phenomenon.

To address the gaps we identified, we recommend focusing qualitative research on answering the following questions: (1) “How does ePROM monitoring impact time use and/or workload?,” (2) “How do different designs of ePROM monitoring frameworks and alert handling mechanisms impact HRU and/or HRQoL?,” and (3) “How/why does ePROM monitoring improve patients’ travel modes and independence or functioning?” Answering these questions through embedding time-motion studies within trials and/or using framework analysis focused on workflow and resource-use themes could help modelers validate model assumptions and more appropriately assign input parameters that match the intervention they wish to simulate.

Equity concerns relevant to ePROMs include digital literacy, limited internet access, disability, language barriers, and limited caregiver support. However, our review has identified limited insights into these issues.

The main exception is the STAR trial, which demonstrated greater benefits from monitoring for patients without prior computer experience [38]. However, the intervention design was altered for those patients to an in-clinic-only schedule, rendering the ePROM collection nonremote for that subgroup.

None of the studies in our review conducted a distributional cost-effectiveness analysis (DCEA). We echo Meunier et al’s [97] recommendations for HTA agencies and governments to publish standardized DCEA methodological guidance and indices to enable future researchers to incorporate routine DCEA implementation in their studies. This would be especially useful in ePROM monitoring economic research.