Advances in public health and health care have substantially improved survival and quality of life across populations [1,2]. Life expectancy has increased in most countries, and many acute conditions that were once fatal can now be effectively treated or managed [2,3]. However, new challenges persist. Mental health problems remain highly prevalent [4], and many people struggle to maintain resilience when faced with major life events and stressful transitions, such as bereavement, unemployment, or divorce [5-8]. Addressing these challenges requires approaches that support both short-term and long-term health and well-being.

Digital health interventions are increasingly positioned as a promising response. They offer flexible, scalable, and cost-effective access to evidence-based support and have been shown to reduce adverse health outcomes, such as stress [9,10], anxiety [11-13], and depression [11-13], in targeted populations, with evidence also supporting benefits for mental well-being at the population level [14]. Yet, most evaluations are based on short-term trials [15-17], often limited to self-reported outcomes [18,19], challenged by retention and adherence [14,17,20], or conducted with relatively small samples (fewer than 200 participants) [21,22].

This has left uncertainty about whether observed effects of digital health interventions persist over time and translate into measurable changes in more objective health outcomes, such as medical prescriptions, primary care usage, or hospitalizations [14,15,18-20]. Consequently, a key gap remains in understanding whether digital health interventions produce long-term health benefits beyond 12 months, particularly when assessed with measures not affected by self-report, attrition, or insufficient statistical power.

To address this gap, we use divorce in Denmark as a case. Divorce is a common life transition in Western societies, with divorce rates ranging from 35% to 50% [9,23]. It is consistently linked to worse health outcomes, including higher risks of chronic disease [24,25], depression [26,27], anxiety [27,28], social isolation [24,29], and reduced physical activity [28,30]. Divorced individuals are also reported to have a 30% higher mortality hazard [8] than their married counterparts, a statistic comparable to other major public health risks, such as physical inactivity and obesity [29]. To examine whether digital health interventions can provide long-term support against these risks, Denmark offers a particularly strong setting: digital literacy is high, uptake of online programs is widespread, and comprehensive national registers enable entire trial cohorts to be followed over time with objective outcome measures [31-33]. These features make divorce in Denmark an ideal case for evaluating the long-term health outcomes of digital health interventions.

The objective of this study is to examine the association between long-term health outcomes and SES One, a digital health intervention developed to support recently divorced individuals in Denmark [9]. SES One is a digital platform accessible online by computer or mobile devices, consisting of learning modules that integrate cognitive behavioral therapy (CBT), narrative therapy, and acceptance and commitment therapy. It was evaluated in a randomized controlled trial (RCT) involving 1856 participants [9]. The RCT demonstrated significant short-term improvements, including reductions in stress, anxiety, depression, and sick days during the first year after divorce [9,34]. Detailed information on the design, methods, analyses, and results of the original RCT can be found in studies by Hald et al [9], Sander et al [34], and Agerholm et al [35].

Building on these findings and on the documented benefits of digital health interventions more broadly, we examine whether the use of SES One is associated with differences in long-term health outcomes. To capture this, we use health care use (including medication use, health care visits, and hospitalizations) as a proxy for health. Health care use is widely applied in longitudinal research, and while short-term help-seeking may support recovery, frequent and persistent use over time has been shown to be associated with prolonged distress and deteriorating physical or mental health [35-39].

We thus link the trial cohort to nationwide registers and follow participants for 5 years. On the basis of the RCT results and existing research in the field, we hypothesize that, compared to controls, participants who had access to SES One will, 5 years after the trial, show (1) reduced mental health medication use (including psycholeptics and antidepressants), (2) less primary care usage (including general practitioner [GP], specialist, and psychologist), and (3) fewer hospitalizations.

Taken together, these outcome measures provide a multidimensional view of somatic and mental health needs and behaviors, as well as provide information on the extent the trial may also lower the health costs associated with divorce. Furthermore, by relying on administrative records, our study is not affected by any differential attrition of respondents in trial follow-up because all trial respondents still alive and in Denmark are consistently included. To our knowledge, this is among the first studies to evaluate long-term (≥5 y) health outcomes of a digital health intervention using objective measures in a large cohort.

SES One (‘SES’ is abbreviated from the Danish ‘Samarbejde Efter Skilsmisse,’ which translates to ‘Cooperation After Divorce’) is a comprehensive, fully digital intervention that comprises 17 modules accessible online via computer or mobile devices, such as smartphones or tablets, and is designed to address key challenges faced by divorced parents. The content is grouped into three main themes that broadly align with evidence-based divorce education priorities identified in the literature: the divorcee, the children, and co-parenting [34]. Each module takes approximately 30 to 60 minutes to complete and incorporates interactive exercises, SMART (specific, measurable, attainable, relevant, and time-bound) goals, and multimedia features, such as voiceovers, animations, expert videos, and dramatized scenarios. The intervention integrates elements from multiple psychological approaches, including CBT, narrative therapy, and acceptance and commitment therapy, and was developed as a “born digital” program specifically designed to leverage digital engagement strategies from the outset. Aside from self-reported survey data, the only additional information collected during the RCT was the number of modules completed. SES One functioned purely as a software-based intervention, without access to health professionals or social workers. Participants could access any module at any time and in any order. During the 12-month RCT period, participants with access to SES One completed an average of 4.27 (SD 2.94) modules [9].

In this retrospective observational study, we observe a treatment group and a control group 5 years after an RCT of SES One. The RCT included 1856 recently divorced Danes who completed an online questionnaire, on average, within a week of their juridical divorce. Study participants were recruited through email invitations distributed by the Danish State Administration, the authority responsible for granting divorce decrees in Denmark [40]. After the online questionnaire, participants were randomly assigned to either the treatment group (n=1031, 55.5%), receiving unlimited access to SES One for 12 months, or the control group (n=825, 44.5%), which did not have access to SES One at any time. The data collection period for the RCT lasted 2 years, from January 2016 to January 2018, with participants invited from January 1, 2016, to January 31, 2017 [9]. For full information and details on the design, methods, analyses, and results of the RCT, please see studies by Hald et al [9], Sander et al [34], and Cipric et al [40].

To conduct the retrospective observational study, we used available Danish register-based data. We linked the deidentified social security numbers of participants in the RCT with Danish national register-based data on the variables of study interest, 5 years after treatment or nontreatment with SES One. As all outcomes were obtained from administrative records, the study did not suffer from bias due to selective attrition in the follow-up period. Only sources of attrition were death and outmigration, which is why any missing information on outcomes was treated as missing at random.

Participants who joined the RCT in early January 2018 but were invited in late 2017 had their treatment date assigned as December 31, 2017 (n=6). This was done to ensure full information on the 5-year follow-up period, which otherwise would not have been possible for this group due to the yearly measurement of primary care usage. We followed STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) reporting guidelines. For the preregistration of this study, please see Hald et al [41].

Statistics Denmark and the ROCKWOOL Foundation have compiled and organized the data. We submitted requests for the specific data of interest and, upon obtaining it, merged the various datasets using social security identifiers (Central Person Register numbers). The requests, organization, and data administration occurred from December 2023 to August 2024, and the final dataset was ready by September 2024. No data linkage of RCT and administrative data occurred before finalizing preregistration on March 15, 2024.

Data on the participants’ social security numbers, treatment status, marriage duration (in months), conflict level, and time of treatment were sourced from the RCT. Data on background variables, including legal gender (man or woman), age (in years), income (total income from all sources before tax in DKK), educational level (highest completed), migration background (Danish or non-Danish origin), and date of juridical divorce, were sourced from Statistics Denmark’s Population Register, Income Register, and Labor Market Register were measured at baseline (juridical divorce).

Data on the outcome variables were sourced from the Danish Patient Register [33], the Danish National Health Service Register [42], and the Medical Prescription Database [43].

The Medical Prescription Database includes all filled outpatient prescriptions in Denmark, allowing us to generate a measure of the total number of filled prescriptions from the date of juridical divorce until 5 years later outside hospital settings, based on Anatomical Therapeutic Chemical (ATC) codes N05 ‘psycholeptics’ (ATC N05: antipsychotics, anxiolytics, hypnotics, and sedatives) and N06A ‘antidepressants’ and N06C psycholeptics and psychoanaleptics in combination. In Denmark, prescriptions of class N06C are not used, so no count of N06C was included in the analyses. Prescriptions are captured at the daily level, allowing us to track individuals from their day of divorce.

Primary care usage represents the total number of billed consultations with general practitioners, specialist practitioners, and publicly funded psychologists from 1 year after the juridical divorce until 5 years later. The data are only available annually and do not report diagnoses nor provide information on date, content, or length of visits. One visit may include more than one billed consultation.

Hospitalizations are measured as a binary variable (yes or no), indicating whether an individual was hospitalized as an inpatient during the period from the date of juridical divorce until 5 years later. The data cover all patients discharged from Danish hospitals. Hospital stays are captured at the daily level, allowing us to track individuals from their day of divorce.

The background and outcome variables have been assessed for all participants included in the study. Descriptive statistics for each background and outcome variable are provided in Table 1, separately for the control group and the treatment group.

Stata/MP 18.0 (Stata Corp) and R 4.5.1 (R Development Core Team) were used for the analyses. The sample consisted of 1856 participants (treatment group, n=1031, 55.5%; and control group, n=825, 44.5%), with the only source of attrition being natural due to outmigration or death (as we expand on this in the Results section). As per preregistration, inferences were made using 2-tailed tests, with P values less than .05 considered statistically significant. For the by-year-since-divorce results, we estimated separate models by year.

To compare the treatment group and the control group on the count variables (medication prescription index and primary care usage index), we used an uncontrolled negative binomial regression model for each outcome. The results are presented as incidence rate ratios (IRRs) interpreted as ratios of expected counts with associated CIs and P values. Moreover, because the primary care usage index is measured annually, we also conducted a sensitivity analysis to examine the effect of starting the count in the year of divorce. These results are provided in the supplementary materials in Multimedia Appendix 1. Finally, especially the medical prescription index has a high share of respondents with a value of zero (n=1342, 72%), we also provide results from a binary logistic regression of any prescriptions filled. Less than 1% of respondents (n=18) had no primary care usage, so we do not include binary regression results for this variable.

To compare the treatment and control groups on hospitalization, we used an uncontrolled logistic regression. We opted for logistic regression over negative binomial regression due to an expected high number of zeros, which was confirmed with 77% of values being zero 5 years after juridical divorce. Results are presented as odds ratios [ORs] with corresponding CIs and P values.

The SES trial participants were recruited from the full universe of Danish divorcees during the trial period, but do not reflect the average divorcee in the period. Previous work has shown that participants were more likely to be women, experiencing their first divorce, and holding longer education [9]. To expand the results beyond the trial participants, we reweight the key findings from the trials to the target population of all Danish residents who finalized a divorce in the same period as the trial took place using calibration weighting [44]. The reweighting is based on a set of covariates that need to be sufficient to capture: “all variables that are related to the trial participation and outcome” [44]. Given that this is a very strong assumption, the results of the calibration weighting should be viewed as providing absolute upper-bound effect estimates under the scenario where, conditional on covariates, the target population would see similar effects of the intervention as the trial population. We limit the results displayed to those found in Tables 2 and 4, and to avoid issues of extreme weights based on few observations, we exclude individuals aged 70 years or higher or who had been married for 45 years or more from both the RCT and the target population (n=57,793).

Table 5 reports the results from the calibration weighting contrasted with the results from the RCT population reported in Tables 2 and 4. Although results are weighted across eight covariates, they differ little from the results reported from the RCT, with the exception of the primary care index, for which the calibration weighted results suggest that an 11.9% decrease in primary care usage (IRR 0.881, 95% CI 0.783-0.996; P=.04; and treatment group average of 45.22 vs 51.30 for the control group) compared to the 5.6% decrease for the RCT sample. Effects on filled prescriptions for psycholeptics were also significant at the 5% level (IRR 0.616, 95% CI 0.395-1.0; P=.048; and treatment group average of 1.211 vs 1.964 for the control group).

The aim of this study was to address the limited evidence on the long-term health outcomes of digital health interventions. By linking a large RCT cohort of recently divorced individuals in Denmark to nationwide registers and following them for 5 years, we were able to assess objective outcomes related to mental health medication use, health care visits, and hospitalizations. The results showed that SES One participants did not have statistically significant lower odds of filling a prescription (OR 0.836, 95% CI 0.681‐1.025; P=.09) but filled 28% fewer prescriptions over the 5-year period (IRR 0.720, 95% CI 0.522-0.993; P=.045) compared to controls, suggesting that the digital health intervention reduced, but not fully removed, the need for medication. This pattern was most evident for psycholeptics and became most pronounced in the fourth year after divorce. No overall differences were found in primary care usage or hospitalizations, but fewer GP visits in years 2 and 3 and fewer hospitalizations in year 4 suggest that some benefits may be late onset.

When the RCT results were reweighted to the full universe of divorcees in Denmark from the same period as the trial ran, the results showed that under the assumption that the covariates included in the recalibration fully accounted for selection into trial participation and the outcome (a strong assumption), the effect of the RCT remained largely unchanged. Given the strict assumption underlying the transportability estimation, this should, however, be viewed as upper bounds on the effect.

Systematic reviews show short-term benefits of digital health interventions for depression, anxiety, and well-being, especially when CBT components and human support are included, but these effects are generally modest and heterogeneous [14,15,19,20]. Studies of fully automated interventions find smaller effects and highlight problems with engagement and adherence [17,23]. Some evidence suggests that resilience- and well-being–focused interventions can sustain benefits up to 12 months, but effects are small [16,18]. Earlier studies on SES One and comparable interventions for divorcees also demonstrate short-term moderate to large health benefits, but these findings are likewise constrained by short follow-up periods or small sample sizes [2,3,43-48].

Longer-term evaluations of digital health interventions are therefore rare, with only a few exceptions. A study on internet-based CBT for social anxiety found durable improvements after 5 years, suggesting that skills learned online can persist without continued contact, although the findings were based on self-report and follow-up participation was selective [24]. Another study on app-supported follow-up after cardiac rehabilitation showed short-term gains that had waned by year 5 once structured support ended [25]. Finally, a 5-year study on a web-based behavior change system for obesity did not sustain weight loss, yet participants started antihypertensive medication less often, pointing to a legacy of early risk factor improvement [49].

Taken together, most evaluations of digital health interventions have been limited by small samples, self-reported outcomes, and follow-up periods of less than 12 months. Against this backdrop, this study advances the field by demonstrating long-term associations between a targeted digital health intervention and objective health outcomes in a large cohort. The findings also open new ways of understanding such interventions by suggesting a “reduce-not-remove” pattern of effects and by pointing to both legacy and late-onset pathways in long-term outcomes.

Interpreting the observed results on mental health prescription use suggests a “reduce-not-remove” effect pattern (ie, dose reduction rather than full discontinuation [45]). The observed difference between the treatment and control groups in the odds of filling a prescription (‘remove’) versus the difference in the expected count of filled prescriptions (‘reduce’) implies that the primary difference between the two groups was a significant and large reduction in the number of prescriptions in the treatment group, not a complete discontinuation of medication. This suggests that the psychosocial tools, strategies, and techniques provided through the intervention helped participants manage stress and emotional distress more effectively [9,28,34], thereby reducing (but not removing) the need for pharmacological treatment. Such partial but meaningful reductions may be a more realistic expectation for the long-term health outcomes of digital health interventions rather than complete discontinuation of treatment.

The observed time-varying associations between treatment and health outcomes point to two possible pathways of long-term health outcomes. One possible pathway is through legacy effects, where the early psychosocial benefits of SES One in the first year after divorce may have helped stabilize participants’ mental health and equipped them with skills to cope with the risks immediately after the divorce. This legacy effect was then sustained over time, contributing to long-term benefits, as seen, for example, in the continued reduction in medication use. The other possible pathway is through late-onset effects, where health benefits only became visible after several years. This interpretation is supported by the observed reductions in GP visits in years 2 and 3 and in hospitalizations in year 4, despite no initial reductions in year 1. In this interpretation, the intervention’s early impact on mental health may have initiated a longer process in which improved psychosocial stability gradually translated into better health outcomes over time compared with controls. For example, non-SES users may have experienced unaddressed psychological distress that accumulated and eventually deteriorated into physical health problems, whereas SES users were protected from such late deterioration due to early stabilization provided by the digital intervention.

Thus, both legacy and late-onset effects may be at play: legacy effects reflecting the initial psychological benefits gained from the intervention and late-onset effects reflecting protection against delayed deterioration.

The findings suggest that digital health interventions such as SES One can provide accessible, nonpharmacological support for individuals navigating major life transitions, such as divorce. By reducing (but not removing) the need for mental health medication and potentially lessening the use of health care services later on, such interventions may complement existing services in primary care and mental health. In practice, this shows the value of integrating digital health into broader care pathways, not as replacements for traditional services, but as scalable tools that can help stabilize well-being and buffer the long-term health risks associated with stressful life events.

Future studies should build on this work by evaluating digital health interventions across different populations, life transitions, and health care contexts to assess generalizability. Longer-term follow-ups using objective measures are essential to understand whether observed benefits are sustained, delayed, or heterogeneous across subgroups. Research should also explore the mechanisms behind legacy and late-onset pathways, including how digital tools interact with health care use, social support, and resilience over time. Finally, comparative studies that include cost-effectiveness analyses could clarify the potential of digital health interventions to complement existing health and welfare systems at scale.

The study also has limitations. First, we could not distinguish between GP visits for mental health versus routine care, which may underestimate SES One’s impact. Second, private psychologist visits and other health services were not tracked, possibly missing some of SES One’s benefits. Third, annual tracking of primary care usage may overlook timing nuances, and measuring only from the posttreatment year may have missed initial effects, while including the treatment year risked pretreatment bias. Fourth, the COVID-19 pandemic, which overlapped with part of the study, could have influenced health care usage, adding or limiting between-group variability. Specifically, individuals may have changed behavior surrounding primary health care usage, which could drive estimates toward zero. Finally, generalizability to other countries may be limited by Denmark’s context, where universal health care and welfare support may have diluted effects [46,47], whereas high digital literacy and advanced digital infrastructure likely strengthened engagement [31-33].

This 5-year follow-up of a postdivorce RCT cohort shows that a digital health intervention was linked to improved long-term health outcomes. The intervention was associated with reduced use of mental health medication over the 5-year period, with reductions in GP visits and hospitalizations observed in specific years. These findings were interpreted as a “reduce-not-remove” pattern and pointed to both legacy and late-onset effects. When findings were transported to the population of all Danish divorces during the trial period using calibration weighting, results remained largely unchanged. Together, the findings indicate that digital health interventions may provide long-term, nonpharmacological support during major life transitions and demonstrate the importance of extended follow-up of digital health interventions, as long-term health outcomes may vary in timing and intensity.