In the past decades, there has been growing evidence for the beneficial effects of contemplative trainings on mental and physical health [1-3] and the development of social skills such as empathy, compassion, and prosocial motivation and behavior [4-6]. However, most of these secular mindfulness or compassion trainings, such as the popular 8-week Mindfulness-Based Stress Reduction program [7], the 8-week Mindful Self-Compassion program [8], or compassion-focused therapy [9], require in-person individual or group sessions.

Recently, and especially in times of the COVID-19 pandemic, considerable awareness has been raised regarding digital mental health solutions to prevailing public health burdens [10,11]. Social emotions such as empathy and compassion have been identified as psychological factors related to mental health resilience [12,13]. In particular, the cultivation of compassion has been called to action to mitigate the impact of the pandemic on psychological well-being [14]. Smartphone-delivered contemplative practices might be promising tools for this endeavor [15,16]. However, such mental trainings are diverse in both the focus and modality of practice. On the one hand, they can focus on cultivating attention-related, socioemotional, or meta-cognitive capacities (for a review of the differential effects of type of practice, see the study by Singer and Engert [1]). In contrast, although most mindfulness-based practices are solitary, more recent approaches have included interpersonal practice [17-19]. The differential effects of the modalities of app-delivered practice and their underlying mechanisms are largely unknown. Therefore, in this randomized controlled trial (RCT), we compared a novel, partner-based socioemotional practice, the so-called Affect Dyad [19], with mindfulness-based practice in its efficacy in increasing empathy and self- or other-related compassion using a multimethod approach. These 10-week trainings were performed in the context of a COVID-19–related mental health study, the CovSocial project [20].

As a social species, humans are equipped with affective and cognitive social capacities that enable an understanding of the feelings, mental states, and intentions of others [21,22]. An important socioaffective capacity is empathy, which describes the ability to share another’s affective state while being aware that its source is the other person [23]. Empathy allows an individual to share both unpleasant and pleasant emotions with others and is ingrained in our neurophysiology [24]. Although empathy helps us better understand one another and foster social cohesion [23], it can also turn into so-called empathic or personal distress when confronted with intense negative feelings or when the self-other distinction fails [25,26]. Empathic distress is a self-related aversive state that is associated with the wish to withdraw and alleviate one’s own negative feelings [25,26]. It presumably is a cause of the observed high burnout rates in health care providers [27,28], a problem that was exacerbated during the COVID-19 pandemic [29]. Compassion and self-compassion have been proposed as protective factors against such negative consequences [26,30].

In contrast to empathy, compassion is characterized by a concern for others, often associated with feelings of warmth and care, and the wish to alleviate their negative feelings [26]. Rather than feeling with, compassion is feeling for a person and, as such, not an emotion per se but a motivation to help others [9,31]. Thus, compassion can be referred to as altruistic other-related capacity and has long been proposed as a crucial driver of prosocial behaviors [32-34]. Indeed, empathic and compassionate responses to others differ on a neurophysiological level, with empathy relying on a brain network associated with pain [24,35] and compassion relying on a brain network associated with positive feelings and reward [26,35,36].

In the past decades, compassion has been conceptually divided into self- and other-related components [8]. Self-compassion has gained extensive attention in resilience research as a means of alleviating distress and promoting well-being [37-39]. The distinction between self- and other-related compassion has shaped contemplative practice and research. Practices such as loving-kindness meditation incorporate successive training of kindness toward the self and others [40,41]. Similarly, questionnaires have been developed to measure both aspects [42,43]. However, most intervention studies to date have focused on either aspect, with a disproportionate focus on self-compassion [44,45]. This study systematically compared 2 different practice types in their effects on both aspects of compassion using a multimethod approach. More specifically, and because of the low reliability of some self-report scales [46], we used a factor analytical approach across multiple scales of compassion for the self and others to derive more reliable higher-level composites of the 2 components of compassion.

A main goal of this RCT was to compare the efficacy of 2 contemplative practice types that differ in content and modality. Generally speaking, mindfulness-based interventions focus on present-moment awareness and attention regulation of thoughts, feelings, or body sensations [47]. With increasing research on trainings that focus on compassion, such as the Mindful Self-Compassion program [8] or compassion-focused therapy [9], debates about the differential efficacy of mindfulness-based and compassion trainings have started to grow.

The findings of a large-scale 9-month mental training study, the ReSource project [41], provide evidence that the type and content of the respective practice matter for a large variety of outcomes [1]. With regard to social emotions, the socioemotional Affect module, which focuses on care, gratefulness, loving-kindness, and acceptance, outperformed the 2 other training modules in increasing other-related compassion in self-report [48] and the computer-based EmpaToM task [49]. Similarly, other studies have shown that, although both mindfulness-based and compassion-based interventions confer benefits on self-compassion, compassion-based approaches have an advantage in promoting compassion toward others [44,50]. In the ReSource project, the sociocognitive Perspective module, with a focus on meta-cognitive abilities and perspective taking, rather than the compassion-based Affect module, could specifically improve theory of mind (ToM) [4]. In contrast to the socioaffective capacities of empathy and compassion, ToM, which is also referred to as perspective-taking or mentalizing [51], is a sociocognitive skill that enables an understanding of others through inference about their thoughts and beliefs [52].

Apart from differences in content (eg, attention, compassion, and meta-cognition), contemplative practices also differ in modality. In contrast to solitary mindfulness-based practice, intersubjective practice formats have recently gained attention [17-19]. Intersubjective mindfulness-based interventions showed additional benefits for self-compassion, empathy, and reduced empathic distress in health care professionals [53]. In dyad practice, 2 individuals are paired in a guided contemplative conversation with switching roles of speaker and listener [19]. Although it has previously been shown that different types of dyads can promote social connectedness and social disclosure [19], the intersubjective format of dyad practice has not yet been compared with solitary practice. In this study, we aimed to test whether (1) low-dose web-based trainings with daily Affect Dyad or mindfulness-based practice could boost social emotions and (2) the Affect Dyad could outperform solitary mindfulness-based practice in improving empathy and compassion because of its explicit focus on cultivating socioaffective capacities such as empathic listening and the acceptance of difficult emotions in the presence of another person.

Finally, our last goal was to gain insights into the psychological processes that underlie the effects of contemplative practices. Despite growing research interest in contemplative trainings, a mechanistic understanding of practice effects is still scarce. According to a current classification system, distinct cognitive mechanisms are relevant to differential practice effects [54]. Thus, the effects of mindfulness-based practice relate to improvements in self-related processes [54-56] such as emotion regulation strategies [57,58], equanimity (ie, a balanced reaction to emotions) [59,60], or embodied cognition and interoceptive awareness [58,61,62]. In particular, the latter is associated with the development of affect-sharing abilities as supported by common neurofunctional correlates [63].

In contrast, loving-kindness and compassion-based practice is proposed to rely on an active generation of positive affect, acceptance, and prosocial motivation [64]. Although the Affect Dyad [19], unlike loving-kindness meditation, does not explicitly train compassion for self and others, it is expected to activate the motivational-biological care system [1,26]. This care system promotes social bonds and feelings of nurture, love, and acceptance [65]. Interestingly, nonjudgmental attitude and acceptance, which are conceptualized as aspects of mindfulness [47], were found to relate to feelings of warmth and care and to be promoted by compassion-based rather than mindfulness-based trainings [48,66,67]. Therefore, we aimed to test whether different psychological processes may underlie the observed training effects.

The first aim of this RCT was to compare the effects of 2 brief app-delivered practice types. Specifically, attention-based, solitary mindfulness-based practices were compared with a novel socioemotional dyadic practice, Affect Dyad [19], on different outcome measures of empathy and compassion for the self and others. Using a multimethod approach, we investigated differential practice effects on validated state and trait questionnaires as well as on an ecologically valid computer task, the EmpaToM [49]. Furthermore, we aimed to investigate the different psychological processes underlying these practice effects. To this effect, in addition to pre- and postintervention assessments, the study design included a weekly assessment of a variety of mediator variables over the 10-week training duration in both intervention groups.

In our preregistered hypotheses [68], we expected that socioemotional dyadic practice would increase compassion significantly more than mindfulness-based practice in both self-report measures and computer-based tasks. We further expected that both trainings would lead to an increase in empathy compared with that in the waitlist control group. The inclusion of sociocognitive ToM performance served as a control condition. As neither practice focused on perspective-taking, no intervention-related change in ToM was expected (see also the study by Trautwein et al [4]).

With regard to the underlying processes of practice effects, we expected that, in the socioemotional training, an increase in compassion would be mediated by socioaffective and motivational processes such as an increase in acceptance and empathic listening and a decrease in empathic distress. Thus, changes in these mediators over time were expected to be significantly associated with changes in compassion from before to after the intervention. We further expected an increase in empathy to be mediated by an increase in mindfulness and interoceptive awareness. Thus, changes in these mediators over time were expected to be significantly associated with changes in empathy from before to after the intervention.

As part of the CovSocial project [20], a longitudinal study on psychological vulnerability, resilience, and social cohesion during the COVID-19 pandemic in 2020 and 2021 in Berlin, 3522 participants from the first project phase, that is, the main study sample in phase 1, were invited to complete a prescreening procedure. With regard to exclusion criteria, participants had to be naive to meditation and yoga practices; have no educational background in psychology; have no current psychopathology, suicidality, chronic illness, or pain; and not use illicit or prescribed substances that affect physiological stress markers.

After the information webinars (see Welcome Days in Multimedia Appendix 1 [19,41]), a senior researcher in the project randomized participants into 3 groups—2 intervention groups and 1 waitlist control group using computer-generated numbers in a lock randomization technique with 1:1:1 allocation. Intervention arms were assigned to participants by the study coordinator. Screening calls were conducted by 4 trained meditation teachers (see Teacher Training in Multimedia Appendix 1) to exclude individuals with clinical levels of psychopathology using the Standardized Assessment of Severity of Personality Disorder [69] and Composite International Diagnostic Screener [70].

A total of 285 participants registered for the RCT, which was in line with the a priori determined sample size (Multimedia Appendix 2 [41,71]). With respect to sociodemographic variables, there was some selection bias toward more female participants and participants with higher levels of education and income compared with the remaining phase 1 participants of the CovSocial project and the Berlin general population (Multimedia Appendix 3 [72]). This study reported on 218 participants (Figure 1 and Table 1). All study participants provided written informed consent. A financial compensation of €10 (US $10.70) per hour was offered.

Self-report and behavioral measures were taken at preintervention assessment and postintervention assessment 1 in the socioemotional, mindfulness-based, and waitlist control groups and at postintervention assessment 2 in the waitlist socioemotional group. In addition, several variables were assessed on a weekly basis in all intervention groups using push notifications on the project’s smartphone app to trace the potential mechanisms of the intervention effects (Figure 2).

This study was conducted in accordance with the Declaration of Helsinki and was approved by the ethics committee of the Charité – Universitätsmedizin Berlin, Germany (EA4/081/21).

Longitudinal changes in acceptance, empathic listening, empathic distress, mindfulness, and interoceptive awareness were modeled separately for the first (socioemotional and mindfulness-based) and second (waitlist socioemotional) intervention phases (Figure 5). In the first intervention phase, acceptance (β_{mindfulness-based}=.04; P=.002) and mindfulness (β_{mindfulness-based}=.02; P=.02) increased significantly in the mindfulness-based group over the course of the 10 weeks of the intervention, whereas empathic distress decreased over time (β_{mindfulness-based}=−.03; P=.009). Empathic listening (β_{mindfulness-based}=.00; P=.48) and interoceptive awareness (β_{mindfulness-based}=.01; P=.11) showed no change during the first intervention period in the mindfulness-based group. In the socioemotional group, there was a significant increase in interoceptive awareness (β_{socioemotional}=.04; P=.003) and mindfulness (β_{socioemotional}=.03; P<.001). No change was found for acceptance (β_{socioemotional}=.02; P=.10), empathic listening (β_{socioemotional}=.01; P=.26), and empathic distress (β_{socioemotional}=−.01; P=.72). None of the interaction terms between intervention group and time reached statistical significance (P>.05). In the waitlist socioemotional group, there was a significant increase in mindfulness (β_{waitlist socioemotional}=.03; P=.002) and interoceptive awareness (β_{waitlist socioemotional}=.03; P=.02) and a trend toward an increase in empathic listening, which was not significant (β_{waitlist socioemotional}=.02; P=.05). No change over time was found in acceptance (β_{waitlist socioemotional}=.02; P=.29) and empathic distress (β_{waitlist socioemotional}=.01; P=.63).

Mediation models showed no (moderated) mediation effects of the slopes of acceptance, empathic listening, empathic distress, mindfulness, or interoceptive awareness on changes from preintervention assessment to postintervention assessment 1 in other-related compassion, behavioral empathy and compassion (EmpaToM), or empathic concern (IRI; Multimedia Appendix 8). In addition, no association between those outcome measures at postintervention assessment 1 and slopes (ie, b-paths) or moderated b-paths by intervention group was found (Multimedia Appendix 9). In the mindfulness-based group, the increase in self-compassion from preintervention assessment to postintervention assessment 1 was found to be mediated by a decrease in empathic distress, with an indirect effect of 0.07 (0.02, 0.14). A decrease in empathic distress could indeed predict levels of self-compassion at postintervention assessment 1 in both the socioemotional and mindfulness-based intervention groups (β=−9.17, 95% CI −15.59 to −3.11). In contrast, an increase in acceptance could predict self-compassion at postintervention assessment 1 only in the socioemotional group (β=6.63, 95% CI 0.52-12.38). In the waitlist socioemotional intervention group, no mediation or regression effects were found for the slopes of acceptance, empathic listening, empathic distress, mindfulness, or interoceptive awareness on changes from postintervention assessment 1 to postintervention assessment 2 (Multimedia Appendices 10 and 11).

The main aim of this RCT was to compare 2 daily app-delivered practices—a solitary mindfulness-based practice and a partner-based socioemotional practice, the Affect Dyad [19]—in their efficacy in boosting empathy and compassion for self and others using a multimethod approach (see Multimedia Appendix 12 for details on standardized reporting of this RCT). The study was embedded in a longitudinal mental health project during the COVID-19 pandemic in a large community sample [20], which was initially randomly recruited through the Berlin registration office.

Indeed, we found evidence that daily 12-minute app-delivered practice over 10 weeks could increase empathy on the EmpaToM task and self- and other-related compassion in a variety of questionnaires. Meanwhile, cognitive perspective-taking abilities (eg, ToM) did not change from before to after the intervention, indicating divergent validity of socioaffective and sociocognitive measures and their differential trainability through contemplative practices. Furthermore, the findings speak to differential intensities in practice effects on self-report and more objective computer task measures, with the latter producing weaker effects. Finally, we found differential psychological processes underlying the 2 interventions, specifically with regard to the development of self-compassion—training-related changes were mediated by a decrease in empathic distress in the mindfulness-based group and were associated with an increase in acceptance in the socioemotional dyad group.

On the basis of a conceptual distinction between compassion toward oneself [39] and compassion as a motivation directed toward the welfare of others [9,26], we used a data-driven strategy to identify whether such a division is empirically supported in the CovSocial community sample of Berliners. Using a factor analytical approach with items from common compassion scales such as the SSCS-L [73], Sussex-Oxford Compassion Scales [43], and FoC [42], indeed 2 factors emerged that can be interpreted as reflective of self- and other-related compassion. The approach of aggregating several compassion scales into 2 distinct composites for self- and other-related compassion led to more reliable outcome measures and further allowed for the testing of potential differential effects of both trainings on different aspects of compassion.

Interestingly, only the subscales of self-kindness and self-judgment of the SSCS-L, a widely used measure of self-compassion, showed relevant loadings on the self-compassion factor. Neff [80] conceptualized self-compassion in 3 dimensions, namely, showing kindness toward oneself, understanding that suffering is part of a shared human experience, and being mindful of one’s own experiences. The 2 subscales integrated into the composite of self-compassion in this study refer to the dimension of self-kindness in a bipolar manner. Although mindfulness is an integral part of the definition of self-compassion by Neff [80], it was not found to be substantially related to the other measures of self-compassion included in this study. Similarly, items that measured the experience of common humanity, which is a part of both the SSCS-L and SOCS-S, did not show relevant loadings on the self-compassion factor, indicating a more distinct aspect of the concept.

Indeed, we observed significant training-related improvements on both aspects of compassion in all 3 intervention groups (socioemotional, mindfulness-based, and waitlist socioemotional), with effects being slightly higher for the self-compassion composite (Cohen d=0.30-0.44) than for the other-related compassion composite (Cohen d=0.24-0.35). Similarly, previous literature reviews and meta-analyses have highlighted that many mindfulness and compassion trainings specifically foster self-compassion [44,50], whereas results on other-related compassion have been less consistent [50]. However, previous studies have disproportionally focused on measuring self-compassion as a training target and less on compassion toward others [44,45]. Thus, based on the inclusion of self- and other-related aspects of compassion in this study, our results more systematically support that, although both aspects can be trained with mindfulness-based as well as dyadic socioemotional practice, there seems to be a slight advantage of such trainings in boosting self-compassion.

On the ecologically valid EmpaToM measures [49], we observed a slightly different pattern of mental training effects. First, and in line with our hypotheses, we found considerable effects only on socioaffective measures (empathy and compassion) but no change in the ToM measure. This is in line with previous findings on practice-specific effects on social skills [4]. Accordingly, perspective-taking can improve through mental practice that is specifically designed to target sociocognitive skills, such as the Perspective Dyad [19]. Furthermore, and again in accordance with our hypotheses, we observed the effects of both interventions on empathy. More specifically, both interventions were found to buffer a general decrease observed in the waitlist control group. The effects on compassion remained below the threshold of statistical significance, but the observed increase was of considerable size for both dyad intervention groups. This replicates previous findings of an increase in compassion on the EmpaToM after 3 months of intense in-person socioemotional but not other types of practice [4]. The small effect sizes of change in task-based compassion (Cohen d=0.08-0.11) highlight the difference in the dose-response relationship on self-report and behavioral measures [81]. Future research including larger sample sizes and longer trainings will have to explore the necessary statistical power and right dose of app-delivered training to bring about robust behavior changes.

When testing the validity of our findings using a frequently used empathy trait measure, the IRI [51], our findings also revealed no significant change in empathic concern, which is in contrast to the increase in the other-related compassion composite. However, similar to the results on the EmpaToM compassion ratings, there was a slight increase in empathic concern after 10 weeks of Affect Dyad training in the waitlist socioemotional group. The lack of change sensitivity of the IRI might be due to its rather trait-like nature (eg, “I would describe myself as a pretty soft-hearted person.”). Furthermore, the empathic concern scale is not per se a direct compassion scale, although it is related in scope. With regard to sociocognitive skills, similar to the results on ToM (EmpaToM) and in line with our hypotheses, the perspective-taking subscale of the IRI showed no training-related change.

Interestingly, both self-report and behavioral measures revealed a general decline in empathy and other-related compassion in the waitlist control group. With preintervention assessments in August 2021 and postintervention assessments in November 2021, this might be reflective of seasonal rhythms in affect that relate to changes in day length [82]. However, self-compassion remained stable in the control group. Alternatively, decreasing levels of empathy and compassion might be explained by an increase in compassion fatigue during the ongoing COVID-19 pandemic [29]. Compassion fatigue is characterized by a reduced capacity for compassion because of extended exposure to others’ negative feelings. However, evidence of affected empathy and compassion during the pandemic is somewhat inconsistent [83].

With regard to the potential underlying psychological processes of the observed practice effects, we indeed found changes in various weekly measured variables during the 10-week intervention periods. Specifically, taking each intervention group alone, mindfulness-based practice significantly decreased empathic distress and increased acceptance and mindfulness over the 10 weeks of training, whereas the partner-based Affect Dyad significantly increased interoceptive awareness and mindfulness in both the socioemotional and waitlist socioemotional groups and increased empathic listening on trend in the waitlist socioemotional group, yet not statistically significant.

Furthermore, mediation analyses revealed that a decrease in empathic distress mediated the observed increase in self-compassion after mindfulness-based training. It has long been argued that mindfulness training can enable individuals to better regulate their emotions and become more equanimous and less reactive to external emotional triggers [60]. Thus, the effects of mindfulness practice on various outcomes might indeed rely on improvements in emotion regulation and other self-related processes [54,58]. In particular, equanimity has been proposed as a key factor in mediating mindfulness effects on compassion by developing a nonjudgmental attitude toward emotions [60]. However, interestingly, we found that decreases in empathic distress specifically mediated changes in self-compassion but not in compassion toward others. Although this result supports a distinction between developing self- and other-related compassion, the generally stronger effects on self-compassion and fatigue effects on other-related compassion in the waitlist control group might also account for it.

In contrast to the effects of mindfulness-based training and our expectations, empathic distress did not decrease during socioemotional dyadic training. A reduction in empathic distress is certainly a desirable training target, particularly in the context of the COVID-19 pandemic [29,30]. As Affect Dyad cultivates empathic listening, sharing difficult emotions may alleviate distress in the speaker but may elicit some empathic distress in the listener, especially in the case of intense emotions [84]. However, note that weekly coaching sessions provided instructions for the listener on how to engage in empathic listening and prevent distress. In line with this, we observed a nonsignificant trend of increased empathic listening in the waitlist socioemotional group and no increase in empathic distress.

However, more importantly, and in line with our hypotheses, we observed that an increase in acceptance could differentially predict levels of self-compassion at postintervention assessment in the socioemotional group compared with the mindfulness-based intervention group. Thus, cultivating acceptance of difficult emotions and gratitude in daily life through the Affect Dyad can foster a care system that relates not only to feelings of nurturing and love for others [65,85] but also to self-love and self-compassion.

One of the limitations of the reported study relates to technical issues that were encountered with the smartphone app that was used for daily practice. During the first few weeks of the first intervention phase, in 25.42% (412/1621) of cases, a stable internet connection could not be established at the first attempt between dyadic partners in the socioemotional training. This might have negatively affected both motivation for and attitude toward daily practice of the Affect Dyad and, thus, caused the higher dropout rate in the socioemotional group (24/95, 25%) than in the mindfulness-based group (16/97, 16%). However, on average, compliance in form of performed daily practice was higher in both socioemotional intervention groups (socioemotional: 88.79%, SD 8.15%; waitlist socioemotional: 90.19%, SD 7.24%) than in the mindfulness-based training (86.4%, SD 14.53%; P<.001). Future studies could help understand whether higher compliance for partner-based over solitary practice is related to social norms given its interpersonal modality.

Despite sample recruitment in accordance with a priori power and sample size calculations, there was a high rate of participant dropout as early as between study onboarding and preintervention assessment (a total of 32/285, 11.2%). As the recruited sample was naive to contemplative practice (note that previous experience was an exclusion criterion), they might have reconsidered participation once introduced to the practice content and requirements of daily training. However, this rather heterogeneous community sample of a COVID-19 mental health study is also an asset with regard to the representativeness of the population. The findings are promising that the 2 app-delivered trainings can indeed be used to boost empathy and compassion in the general population.

Previous research has suggested an advantage of socioemotional contemplative training, including loving-kindness meditation, on compassion [4,48]. However, these findings were based on an extensive training study that combined several training approaches [41]. For the first time, this RCT selectively probed the efficacy of the app-delivered partner-based Affect Dyad in comparison with mindfulness-based practice. The novelty of this study also relates to the modality of the Affect Dyad. Thus, findings are still scarce on positive effects of partner-based contemplative practice on socioaffective capacities [53]. Indeed, the 10-week low-dose web-based practice with the Affect Dyad in this study could replicate the effects of the 3-month Affect module in the ReSource project in increasing empathy and compassion [4,48].

In this RCT, we showed that app-delivered, mindfulness-based, and partner-based socioemotional practices can promote empathy and compassion for self and others. Thereby, the effects were stronger for state self-report measures than for task-based or dispositional measures, pointing to a lower dose-response relationship in those markers. In addition, we found differential mechanisms for practice effects on self-compassion. Although mindfulness practice takes effect via a decrease in empathic distress, changes in self-compassion after Affect Dyad training were associated with an increase in acceptance, which might be related to a more general activation of a motivational human care system. In sum, the findings support that app-delivered trainings such as these could be a promising scalable solution to boost social skills and, thus, mitigate the threats and challenges to social cohesion, mental health, and psychological well-being faced by societies worldwide.