Published on in Vol 24, No 11 (2022): November

Preprints (earlier versions) of this paper are available at, first published .
Personalized Prediction of Response to Smartphone-Delivered Meditation Training: Randomized Controlled Trial

Personalized Prediction of Response to Smartphone-Delivered Meditation Training: Randomized Controlled Trial

Personalized Prediction of Response to Smartphone-Delivered Meditation Training: Randomized Controlled Trial

Original Papetar

1Harvard Medical School, Boston, MA, United States

2McLean Hospital, Belmont, MA, United States

3Center for Healthy Minds, University of Wisconsin – Madison, Madison, WI, United States

4Department of Psychology, University of Wisconsin – Madison, Madison, WI, United States

5Department of Psychiatry, University of Wisconsin – Madison, Madison, WI, United States

6Department of Counseling Psychology, University of Wisconsin – Madison, Madison, WI, United States

Corresponding Author:

Simon B Goldberg, PhD

Department of Counseling Psychology

University of Wisconsin – Madison

315 Education Building

1000 Bascom Mall

Madison, WI, 53706

United States

Phone: 1 608 265 8986


Background: Meditation apps have surged in popularity in recent years, with an increasing number of individuals turning to these apps to cope with stress, including during the COVID-19 pandemic. Meditation apps are the most commonly used mental health apps for depression and anxiety. However, little is known about who is well suited to these apps.

Objective: This study aimed to develop and test a data-driven algorithm to predict which individuals are most likely to benefit from app-based meditation training.

Methods: Using randomized controlled trial data comparing a 4-week meditation app (Healthy Minds Program [HMP]) with an assessment-only control condition in school system employees (n=662), we developed an algorithm to predict who is most likely to benefit from HMP. Baseline clinical and demographic characteristics were submitted to a machine learning model to develop a “Personalized Advantage Index” (PAI) reflecting an individual’s expected reduction in distress (primary outcome) from HMP versus control.

Results: A significant group × PAI interaction emerged (t658=3.30; P=.001), indicating that PAI scores moderated group differences in outcomes. A regression model that included repetitive negative thinking as the sole baseline predictor performed comparably well. Finally, we demonstrate the translation of a predictive model into personalized recommendations of expected benefit.

Conclusions: Overall, the results revealed the potential of a data-driven algorithm to inform which individuals are most likely to benefit from a meditation app. Such an algorithm could be used to objectively communicate expected benefits to individuals, allowing them to make more informed decisions about whether a meditation app is appropriate for them.

Trial Registration: NCT04426318;

J Med Internet Res 2022;24(11):e41566




Precision medicine, which involves the use of individual variability to guide prevention and treatment, has gained momentum in the health sciences over the past several years [1]. This approach aims to improve outcomes by matching patients with interventions most likely to yield success. In some medical specialties, precision medicine has led to impressive advances in personalized care. For example, research in oncology (eg, lung and breast cancer) has effectively matched patients to targeted cancer treatments based on the unique genetic characteristics of their tumors, which has been shown to improve outcomes [2-4].

Psychiatry and clinical psychology have long hoped to better match patients with interventions. Numerous studies have examined patient-level factors (eg, demographic, clinical, and neurobiological characteristics) as predictors of treatment response [5,6]. However, with many potential predictors and inconsistencies across studies in the presence, direction, and strength of associations with outcomes, empirically supported guidelines for optimal treatment matching remain elusive.

Machine learning has emerged as a promising analytical approach well suited for handling and integrating large numbers of predictor variables, including correlated predictors, that may individually only modestly predict outcomes of interest but can collectively predict significant variance in patient outcomes [7,8]. Specific machine learning approaches such as decision tree–based algorithms (eg, random forest) also effectively model nonlinear and higher-order interactions that may underlie predictive relationships [9]. In contrast to traditional statistical approaches that emphasize evaluating a specific hypothesis (ie, null hypothesis significance testing), machine learning models typically emphasize optimizing predictive performance, and evaluating the generalizability of models to new individuals (eg, via cross-validation [CV], hold-out samples, or external validation) [10]. Machine learning approaches are increasingly being applied with some success in psychiatry and clinical psychology, with a growing number of studies demonstrating their ability to predict response to various psychiatric treatments [10-12].

In pursuit of precision mental health, researchers have leveraged machine learning approaches to optimize treatment recommendations [13-15]. For example, DeRubeis et al [16] developed the Personalized Advantage Index (PAI) as an algorithm for guiding treatment recommendations based on pretreatment patient characteristics. These models attempt to predict the benefit that a specific patient would derive from treatment A versus treatment B. The PAI has been successfully used to predict response to cognitive behavioral therapy (CBT) versus an antidepressant medication [16], CBT versus interpersonal therapy [17], CBT versus psychodynamic therapy [18], and an antidepressant medication versus placebo [19].

Prior research using the PAI and related approaches [12] provides promising initial evidence that data-driven algorithms may improve patient outcomes by matching individuals to the most therapeutically beneficial treatment, as opposed to the current suboptimal trial and error approach to treatment selection, which results in protracted psychiatric illness until an effective treatment is found. However, the fact remains that a substantial proportion of individuals with psychiatric disorders go untreated [20,21]. Digital health technologies, such as internet-based CBT [22] and smartphone-delivered mental health apps [23], have the potential to substantially increase access to evidence-based treatments [24]. However, the availability of thousands of mental health apps leaves potential consumers faced with a dizzying number of choices, with essentially no way of knowing which specific app may best suit their needs [25]. Data-driven treatment recommendation algorithms, such as the PAI, offer promising tools for informing optimal patient-treatment fit. Such approaches may also be valuable for addressing the persistent limitations of mobile health (mHealth) approaches, including notoriously high and rapid disengagement [26,27]. Moreover, the scalability of mHealth makes it possible to collect adequately powered sample sizes for robust modeling [28].

A recent analysis of available mental health apps revealed that meditation and mindfulness training (along with journaling and mood tracking) are the most common features offered across apps [29]. The two most widely used meditation apps (Headspace and Calm, with 5 million and 9 million monthly active users, respectively) account for 96% of daily active users in a recent evaluation of the top 27 apps for depression and anxiety [30]. Despite the soaring popularity of meditation apps, a critical question remains unanswered: For whom is app-based meditation training well suited?

This Study

This study involved a secondary analysis of a large-scale randomized controlled trial (RCT) comparing a meditation-based smartphone app, the Healthy Minds Program (HMP), with an assessment-only control condition [31]. The RCT was conducted on a sample of school district employees (n=662) in the state of Wisconsin during the COVID-19 pandemic. Relative to prepandemic levels, the rates of emotional distress and depressive symptoms increased substantially during the COVID-19 pandemic [32]. Available evidence suggests that the emotional well-being of teachers also decreased during the pandemic [33,34], as they coped with COVID-19–related stressors, uncertainty, and risks with the return to in-person instruction. Using data from the above RCT, the overarching goal of this study was to develop and evaluate a data-driven (PAI) approach to inform personalized meditation app recommendations for school employees. Using readily gathered self-reported baseline demographic and clinical characteristics, we developed and tested a machine learning algorithm to identify which individuals are most likely to benefit from the HMP app.

Participants and Procedure

Wisconsin school district employees were recruited via email and other electronic media between mid-June 2020 and late August 2020 (for a full description of study procedures, refer to the study by Hirshberg et al [31]). Eligible participants were adults (aged ≥18 years) currently employed by a Wisconsin school who owned a smartphone capable of downloading the HMP, were fluent in English, had limited exposure to meditation or the HMP app, and had depressive symptoms below the severe range (t score<70 on Patient-Reported Outcomes Information System [PROMIS] Depression [35]). The t scores are population normed at 50, with an SD of 10. On completing the pretest measures, 666 participants were randomly assigned to use the 4-week HMP or an assessment-only control condition (4 participants were removed for failing multiple attention checks; refer to Figure S2 in Multimedia Appendix 1 [31,35-49] for the CONSORT [Consolidated Standards of Reporting Trials] flow diagram). Participants completed weekly questionnaires during the intervention period (ie, weeks 1, 2, and 3) along with a posttreatment assessment (week 4) and follow-up assessment (3 months after the end of the intervention period). These measures were administered via the web-based REDCap (Research Electronic Data Capture) survey system.

The trial was preregistered at (NCT04426318) and through the Open Science Framework [50]. However, the current prediction analyses were not planned a priori and were not included in the preregistered data analysis plan. All code (implemented in the R statistical software [51]) used to reproduce the analyses in the manuscript have been posted on Open Science Framework [52].

The HMP includes contemplative practices designed to build skills supportive of 4 pillars of well-being: awareness, connection, insight, and purpose [36,37]. Participants were encouraged to engage with content from each of the 4 modules for approximately 1 week (ie, 4 weeks total). The content included didactic instruction and guided meditation practices. For the guided practices, participants could select the length of practice from 5 to 30 minutes. The HMP app was used for a mean of 10.9 (SD 9) days during the 4-week trial. For additional trial and sample details, refer to the study by Hirshberg et al [31].

Ethics Approval

The study procedures were approved by the University of Wisconsin—Madison Institutional Review Board (number 2020-0533).


Demographic Characteristics

The participants reported their age, gender identity, race and ethnicity, marital status, and income at baseline.

Primary Outcome

The prespecified primary outcome in the parent RCT was psychological distress, which was a composite of the computer-adaptive versions of the PROMIS Anxiety and PROMIS Depression measures [35] and the 10-item Perceived Stress Scale [38]. All 3 are widely used measures with established reliability and validity [39,40]. Refer to Multimedia Appendix 1 for details. Consistent with the prespecified data analytic plan, multilevel models estimated changes in distress over the 4-week intervention period. Random slopes (representing individual changes in distress over the intervention period) were calculated for each participant and served as the primary outcome in our machine learning prediction models.


Several additional self-report questionnaires assessed secondary outcomes and candidate mediators that were theoretically linked to the pillars of well-being trained within the HMP. The 15-item Perseverative Thinking Questionnaire (PTQ) [53] assessed worry and rumination (Cronbach α=.95). The 5-item World Health Organization [54] assessed global well-being (α=.85). The 8-item Act with Awareness subscale of the Five Facet Mindfulness Questionnaire [55] assessed mindful attention in daily life (α=.91). The 5-item National Institutes of Health Toolbox Loneliness Questionnaire [56] assessed perceived social disconnection (α=.90). The 12-item Self-Compassion Scale Short Form [57] assessed feelings of kindness toward oneself (α=0.86). The 10-item Drexel Defusion Scale [58] assessed the ability to experientially distance from internal experiences (α=.84). The 10-item Meaning in Life Questionnaire (MLQ [59]) assessed the presence and search for meaning (Cronbach α=.91 and Cronbach α=.93, respectively).

Analytic Strategy

Predictor variables included preintervention distress (composite measure), anxiety (PROMIS), depression (PROMIS), stress (Perceived Stress Scale), repetitive negative thinking (PTQ), the mindfulness facet of acting with awareness (Five Facet Mindfulness Questionnaire), loneliness (National Institutes of Health Toolbox Loneliness), defusion (Drexel Defusion Scale), presence (MLQ), search for meaning (MLQ), self-compassion (Self-Compassion Scale Short Form), well-being (5-item World Health Organization), age, gender, race, marital status, and income.

Missing Value Imputation

Missing data were imputed using a random forest–based imputation (MissForest package in R [60]). To avoid contamination between the predictor and outcome scores, which may optimistically bias predictive performance, the outcome variable (slope of change in distress) was excluded from the imputation procedure. The rate of missing data was very low, with no variable missing more than 6 values. Refer to Multimedia Appendix 1 for additional details.

Generating Predicted Outcomes

To predict outcomes, 2 prognostic models (using elastic net regularized regression [ENR]; glmnet package in R) were developed: one for participants who received HMP and one for those who received the assessment-only control condition. To minimize overfitting, which can occur with traditional k-fold CV, a nested CV procedure was used for each of these prognostic models (ie, incorporating an outer and inner CV loop [41-44]). Refer to Multimedia Appendix 1 for details of the nested CV procedure.

The steps mentioned earlier generated predicted HMP outcomes for HMP participants and predicted control condition outcomes for the control participants. To generate predicted outcomes for the counterfactual condition (ie, the treatment condition one did not receive), an ENR model was developed for one group (ie, full HMP or control sample) and used to predict outcomes for participants in the other group.

Evaluation of Recommendations

As a final product of the prediction models mentioned earlier, every participant had 2 predicted outcome scores: one for HMP and one for the control condition. Consistent with previous similar studies [18,19,61], we computed a PAI score by subtracting these 2 predicted outcomes (ie, the predicted slope of change in distress for HMP minus control) for each individual. Thus, a negative PAI score indicates that a given participant is predicted to experience greater reductions (ie, a more negative slope) in distress in HMP relative to the assessment-only control condition (and vice versa for positive PAI scores). The PAI can be interpreted as a continuous indicator reflecting the expected magnitude of the advantage of one treatment condition relative to the other (eg, a large negative PAI value indicates that the model predicts a relatively large between-group difference in outcome favoring HMP). We tested whether PAI scores moderated treatment group differences in outcome (ie, slope of change in distress) via a group (ie, intervention condition) × PAI interaction. The latter test allowed us to answer the following question: Are more negative PAI scores (reflecting relatively greater predicted benefit from HMP relative to the control condition) in fact associated with larger observed differences in outcomes favoring HMP?

Comparison Model

We compared the abovementioned multivariable machine learning (ENR) model with a simple linear regression with baseline repetitive negative thinking (PTQ) scores as the sole predictor (ie, repeating the above steps to generate a PAI score for every participant) implemented via 10-fold CV (repeated 100 times to generate stable estimates). Repetitive negative thinking was selected as a predictor in this comparison model based on prior research, indicating that it predicts response to mindfulness apps [43,45]. Refer to Multimedia Appendix 1 for additional analyses with baseline distress as the sole predictor. Finally, we used the parameter estimates from the final models to demonstrate the translation of the predicted outcomes to personalized recommendations for app-based mindfulness training.

All analyses were conducted using R software (version 4.0.2) [62]. The sample size was originally determined for the purpose of the parent trial to detect between-group differences in the primary outcome (change in distress [50]). To estimate whether the current sample size was adequately powered for the analyses proposed in this study, a Monte Carlo simulation approach (InteractionPoweR package in R) was used. Informed by the effect sizes from a prior mindfulness app RCT [45] that tested similar group × PAI interactions, simulations revealed that a sample size of at least 153 was needed for group × PAI interaction tests (with Cronbach α=.05; power=80%; refer to Figure S1, including figure note, in Multimedia Appendix 1 for additional power analysis details).

Sample Demographics

The majority (523/662, 79%) of the participants reported depression or anxiety symptoms at baseline that were above the clinical cutoff on the PROMIS Depression and PROMIS Anxiety measures (t score>55), and more than half of the sample (343/662, 51.8%) reported moderate or greater anxiety or depressive symptoms at baseline (t score>60).

The groups did not differ at baseline in terms of the demographic or clinical variables (Table 1). Of those assigned to HMP, 95.6% (329/344) downloaded the app and 78.8% (271/344) used the app for ≥1 day. The mean number of days of use was 10.88 (SD 9.08). The mean number of minutes of practice was 127.93 (SD 130.63).

Table 1. Descriptive statistics for Healthy Minds Program and assessment-only control at baseline.
VariableHealthy Minds ProgramControlPa value

Value, NValue, n (%)Value, mean (SD)Value, NValue, n (%)Value, mean (SD)
Age (years)34442.47 (11.06)31842.70 (10.23).78
Gender (female)344299 (86.9)318279 (87.7).75
Non-Hispanic White344304 (88.4)318268 (84.3).13
Married344243 (70.6)318216 (67.9).45
College education343308 (89.8)316281 (88.9).72
Income (US $)

≤50,00034456 (16.3)31855 (17.3).73

50,000-100,000344141 (41.0)318129 (40.6).91

100,000-150,000344104 (30.2)31896 (30.2).99

≥150,00034440 (11.6)31837 (11.6).99

Depression34255.37 (6.20)31555.47 (6.43).85

Anxiety34259.83 (6.95)31560.00 (7.11).75
Perceived Stress Scale3422.89 (0.56)3152.87 (0.60).69
Distressc (composite)3420.00 (0.88)3150.00 (0.91).97
Perseverative Thinking Questionnaire34229.89 (10.43)31529.62 (11.29).76
Five Facet Mindfulness Questionnaire—Acting with Awareness Subscale34224.80 (5.93)31524.56 (6.12).62
National Institutes of Health Toolbox Loneliness3422.53 (0.77)3152.58 (0.77).45
Drexel Defusion Scale34224.83 (7.89)31524.50 (8.16).60

Presence34226.20 (5.44)31525.81 (5.46).36

Search for meaning34221.63 (6.61)31522.09 (6.79).38
World Health Organization well-being34112.76 (4.71)31512.47 (4.33).42
Self-Compassion Scale3422.98 (0.69)3152.93 (0.70).37

aP value from independent samples t test comparing groups at baseline.

bPROMIS: Patient-Reported Outcomes Information System.

cDistress: composite of PROMIS Depression, PROMIS Anxiety, and Perceived Stress Scale.

dMLQ: Meaning in Life Questionnaire.

Outcome Prediction

Higher baseline levels of distress, depression, and stress predicted better outcomes (ie, greater reductions in distress) in HMP (Table 2). The zero-order correlations between outcome and these 3 predictors were r=−0.30 (for distress), r=−0.30 (depression), and r=−0.26 (stress). Predicted HMP outcomes were significantly correlated with observed outcomes for the HMP group (r=0.27; P<.001; root mean square error [RMSE]=0.10) but not with the control condition outcomes (r=0.07; P=.21; RMSE=0.12). Conversely, predicted control condition outcomes were significantly correlated with observed outcomes for the control group (r=0.19; P<.001; RMSE=0.10) but not with HMP outcomes (r=0.10; P=.06; RMSE=0.12). Higher baseline scores for the following variables predicted better outcomes in the control condition: distress, anxiety, depression, stress, loneliness, defusion, and presence. In addition, lower levels of repetitive negative thinking, higher self-compassion, and being married were associated with better control condition outcomes (Table 2).

Table 2. Baseline variables retained in elastic net models predicting outcomes for each conditiona.
PredictorsHealthy Minds Program model, coefficientControl model, coefficient
Age (years)b
Marital status−0.006


Perceived Stress Scale−0.003−0.006
Distressd (composite)−0.011−0.008
Perseverative Thinking Questionnaire0.012
Five Facet Mindfulness Questionnaire—Acting with Awareness Subscale
National Institutes of Health Toolbox Loneliness−0.002
Drexel Defusion Scale−0.011


Search for meaning
World Health Organization well-being
Self-Compassion Scale−0.002

aThe larger set of baseline predictors retained in the elastic net regularized regression model applied to the control participants relative to the Healthy Minds Program (HMP) group was because the best-fitting model in the former group had a lower α value (ie, closer to ridge than lasso regression) relative to the HMP group. Negative parameter estimates indicate that higher scores on the predictor variable are associated with better outcomes (ie, reductions in distress).

bVariables that were not retained in the elastic net model.

cPROMIS: Patient-Reported Outcomes Information System.

dDistress: composite of PROMIS Depression, PROMIS Anxiety, and Perceived Stress Scale.

eMLQ: Meaning in Life Questionnaire.

Meditation App Recommendations

The mean PAI score was −0.07 (SD 0.03; range −0.17 to 0.03), indicating that the model predicted greater average symptom improvement for the HMP meditation app than for the assessment-only control condition. The model recommended HMP (PAI<0) for all participants except 5 (657/662, 99.2%).

Evaluation of Recommendations

A significant group × PAI interaction emerged in predicting outcome (t658=3.30; P=.001; adjusted r2=0.10), indicating that PAI scores moderated group differences in outcome. As displayed in Figure 1, as PAI scores decrease (ie, reflecting relatively stronger HMP recommendations), group differences in observed outcome increase, favoring HMP.

Figure 1. Group × Personalized Advantage Index (PAI) interaction. As PAI scores decrease (ie, reflecting relatively stronger recommendations for the Healthy Minds Program [HMP] app) group differences in observed outcome increase, favoring HMP.
View this figure

Comparison Model

In the linear regression comparison model applied to the HMP group, higher levels of repetitive negative thinking were significantly associated with a greater reduction in distress from the mindfulness app (B=−0.02; t342=−3.37; P<.001). The correlation between predicted HMP outcomes and observed outcomes was r=0.16 (P=.003; RMSE=0.10) for participants who received HMP and r=−0.14 (P=.02; RMSE=0.12) for the control group. In contrast to the pattern of findings for the HMP group, the linear regression model applied to the control sample revealed that higher levels of repetitive negative thinking were significantly associated with poorer outcomes than in the control condition (B=0.01; t316=2.44; P=.02).

The correlation between predicted control condition outcomes and observed outcomes was r=0.11 (P=.049; RMSE=0.11) for the control group and r=−0.18 (P<.001; RMSE=0.12) for the HMP group.

A significant group × PAI interaction emerged in predicting changes in distress (t658=3.81; P<.001; adjusted r2=0.11), indicating that PAI scores moderated group differences in outcomes (Figure 2). Specifically, as PAI scores decreased (reflecting increasing repetitive negative thinking scores), group differences favoring the HMP condition also increased. Given the association between repetitive negative thinking and depressive symptoms [46,47], we also conducted additional sensitivity analyses controlling for baseline levels of depressive symptoms (as well as considering the number of days the app was used), which yielded the same pattern of findings (Multimedia Appendix 1). In summary, these results indicate that a simple linear regression including repetitive negative thinking as the sole predictor yields equivalent performance relative to a more complex multivariable ENR model (ie, adjusted r2=0.11 vs r2=0.10, respectively, for the group × PAI interaction).

Figure 2. Group × Personalized Advantage Index (PAI) interaction for the comparison model (ie, linear regression with baseline repetitive negative thinking [PTQ] scores as the sole predictor). As PAI scores decrease (ie, reflecting relatively stronger recommendations for the Healthy Minds Program [HMP] app) group differences in observed outcome increase, favoring HMP.
View this figure

Translating a Predictive Model to Personalized Meditation App Recommendations

To demonstrate the translation of a predictive model to personalized recommendations, we used the parameter estimates from the above regression models to estimate predicted changes in distress in HMP versus the assessment-only condition for a new individual based on their preintervention repetitive negative thinking score. Given that the simpler regression model performed similarly to the more complex multivariable ENR models, we used the former model for this demonstration.

First, as shown in Figure 3, we plotted the relationship between PAI scores and outcomes for HMP (blue line) and the assessment-only control condition (red line). The dashed vertical gray line represents the point at which the 2 regression lines intersect. An individual with a PAI score to the left of this line was predicted to have a better outcome in HMP relative to the assessment-only control condition (and vice versa for individuals with PAI scores to the right of this line). The area to the left of this line is colored yellow, reflecting a “cautious recommendation” for app-based meditation training. Second, we computed a 95% CI via bootstrap resampling (Boot package in R) [63]. Specifically, we drew 1000 samples with replacement and recomputed the 2 regression lines and their intersection points in each of these samples. The dashed vertical red line represents the left margin of the 95% CI for this intersection point. In other words, if an individual’s PAI score falls to the left of this line, our confidence in the predicted benefit of HMP relative to the assessment-only condition increases. Third, we also implemented the Johnson-Neyman technique [64] (Interactions package in R) to probe the group × PAI interaction and to estimate the value of the moderator (PAI) at which group differences in outcomes become statistically significant. This occurred at PAI<−0.02 (solid vertical gray line in Figure 3, immediately adjacent to the dashed red line). If a participant’s PAI score falls to the left of both the 95% CI (dashed red line) and the Johnson-Neyman threshold (solid gray line), the plot area is colored green to reflect a more confident recommendation to use HMP.

To illustrate with a concrete example, an individual with a repetitive negative thinking (PTQ) score of 1 SD above the mean (ie, 41) would have a PAI score of −0.10 (within the “green zone” of Figure 3) and a predicted slope of change in distress of −0.049 (ie, expected reduction in distress) in HMP and 0.047 (ie, expected increase in distress) in the assessment-only condition over 4 weeks. Assuming that this individual had a preintervention level of distress at the 50th percentile, they would be predicted to be at the 41st percentile (relative to preintervention distress scores) following the 4-week mindfulness app course and at the 58th percentile if they only completed symptom assessments (ie, control condition). In summary, based on a brief assessment of perseverative negative thinking, our algorithm can provide individual users with useful information regarding their expected benefit before they decide to enroll in a multiweek course of app-based meditation training.

Figure 3. Plot of the relationship between Personalized Advantage Index (PAI) scores and outcome for each condition to inform personalized recommendations. The dashed vertical gray line indicates the point at which the 2 regression lines intersect (left margin of a bootstrapped 95% CI is shown with a dashed vertical red line). The solid vertical gray line (adjacent to the red line) is derived from the Johnson-Neyman technique and represents the value of the moderator (PAI) at which between-group differences in outcome become statistically significant. Refer to the detailed description in text, with an example for personalized Healthy Minds Program [HMP] recommendation.
View this figure

Principal Findings

An increasing number of individuals are turning to meditation apps to alleviate their emotional distress. Meditation apps represent the most commonly used mental health apps for depression and anxiety [30]. Despite their growing popularity, little is known about the benefits of these apps. In this study, we developed an algorithm to predict the benefit that an individual would be expected to experience from a smartphone-based meditation intervention (HMP) relative to an assessment-only control condition. We found evidence that a data-driven model can successfully predict differential response to a meditation app versus an assessment-only control condition using self-reported baseline demographic and clinical characteristics. Specifically, PAI scores significantly moderated group differences in outcomes. Individuals with more negative PAI scores, reflecting relatively stronger meditation app (ie, HMP) recommendations, had better outcomes if randomly assigned to the meditation app relative to the control condition. As expected, given overall group (ie, HMP > control) differences in outcome [31], the models typically predicted greater benefits from HMP versus the control condition. However, the predicted benefits of HMP were not always large, and in some cases, the PAI model predicted either relatively small between-group differences in outcome (“yellow zone” in Figure 3) or even better outcomes in the control condition (“red zone”). The former cases could be interpreted as instances in which the costs of engaging in a multiweek meditation app course (eg, time investment, delay in engaging with other, more helpful interventions) may not be worth the potential benefits.

Critically, a comparison linear regression model that only included information about baseline levels of repetitive negative thinking performed comparably well to a multivariable machine learning model (in contrast, refer to the studies by Webb et al [65] and Buckman et al [66]). Repetitive negative thinking moderated the outcome of app-based meditation training relative to the assessment-only control condition. Importantly, these findings reveal that higher repetitive negative thinking is not simply a general “prognostic” indicator of one’s likelihood of experiencing reductions in distress (eg, due to regression to the mean or the passage of time). In other words, greater repetitive negative thinking did not predict greater reductions in distress in both the meditation app and control conditions. Instead, and similar to prior research focused on a different mindfulness app and sample (adolescents with elevated rumination) [43,45], individuals with higher baseline levels of repetitive negative thinking derived greater relative benefit from a meditation app. One question is whether these findings are specific to repetitive negative thinking or instead may be driven by correlated clinical characteristics, in particular, depressive symptoms or distress. Sensitivity analyses revealed that repetitive negative thinking significantly moderated group differences in outcomes even when controlling for depressive symptom severity or distress (Multimedia Appendix 1). In summary, these findings indicate that a brief self-report assessment of repetitive negative thinking can inform which individuals are most likely to benefit from app-based meditation training.

As illustrated in Figure 3, our predictive model can be readily applied for personalized meditation app recommendations for new individuals. First, the model provides a binary prediction of whether an individual is expected to experience greater reductions in distress from the meditation app relative to symptom assessment only (ie, based on whether PAI scores fall to the left or right of the intersection point [vertical dashed gray line]). Second, the model provides an estimate of the magnitude of the expected difference in outcomes between the meditation app and the control condition. Finally, the model also distinguishes between the strengths of recommendations to use the meditation app, demarcated by the green (confident recommendation) and yellow (cautious recommendations) zones of the figure (with boundaries defined by a bootstrapped CI and Johnson-Neyman interval). Collectively, this information can be used to provide individuals with objective metrics about expected outcomes to inform their decision on whether to enroll in a meditation app course. Such information could readily be implemented within mHealth interventions such as the HMP. Participants could first complete a brief self-report assessment of repetitive negative thinking and receive feedback on their predicted outcomes before deciding to use the app.

Although potentially useful in terms of encouraging the optimal use of users’ time and attention, informing some individuals that engagement with a meditation app may not be beneficial to them is unlikely to be embraced by many intervention developers. However, these models can be readily extended to instances in which one or more mHealth interventions are being compared. Given the thousands of available mental health apps [25], which should be compared? One approach is to focus on the most popular (eg, most frequently downloaded) mental health apps, which include mindfulness, journaling, CBT, and mood tracking apps [29,30]. For example, future studies could develop algorithms for predicting response to various popular mental health apps, which differ substantially in intervention focus (eg, meditation apps vs CBT-based apps vs mood tracking) [29,67], or even compare a mental health app with conventional (in-person) psychotherapy or pharmacotherapy. Such studies could determine, for example, whether we can predict which individuals with depressive symptoms require conventional, face-to-face CBT (or an antidepressant prescription) versus those who would experience symptom remission from a brief app-based meditation or CBT course. In addition, future studies could compare different versions of a single app. For example, individuals may differ in the extent to which they benefit from different types of meditation (eg, cultivating focused attention on breath, open monitoring, or loving-kindness meditations) or different lengths or frequencies of guided meditation sessions.

In addition to informing consumer choice, the ability to predict who is most likely to benefit from a particular intervention could inform health care policy and decision-making. In contrast to a stepped care model in which treatment intensity is escalated based on the response to interventions, predictive models could be used to initially assign patients to the treatment expected to yield the best outcomes for that individual based on their baseline characteristics (ie, stratified care) [68]. In theory, the latter approach could minimize the delay in receiving an effective intervention.

Another important avenue for future research is to test the extent to which these findings can be generalized to other meditation apps (eg, Headspace and Calm). In many ways, HMP is similar to other meditation apps. It includes training in mindfulness and connection (eg, loving-kindness, compassion) practices that are also available in popular mindfulness apps such as Headspace and Calm. One difference is that HMP includes practices designed specifically to cultivate a healthy sense of self (Insight module) as well as purposes and meaning in life (Purpose module). The inclusion of these practices is derived from a neuroscience-based model of well-being on which HMP is based [36]. Thus, it is more accurate to view HMP as a meditation app that intentionally moves beyond mindfulness to place equal emphasis on other domains of well-being and contemplative practices designed to support these additional domains. Ultimately, additional research is needed to test whether the pattern of findings presented in this study generalize to other meditation and mindfulness apps.

Finally, given the lack of prior research predicting mental health app outcomes, further research is needed to test the impact of presenting predicted mindfulness app prognosis on patient outcomes. For example, before using a mindfulness app, patients could be randomly assigned to receive their predicted outcomes or not receive this information. Several relevant outcomes could be examined, including (1) between-group differences in symptom change, (2) the extent to which receiving these predictions influences expectancies of therapeutic benefit, (3) the relationship between expectancies and app outcome, and (4) the extent to which individuals use the algorithm-recommended intervention or disregard the recommendation.


This study has several important limitations. First, although basing models exclusively on self-reported data is attractive from an implementation perspective, we may have excluded other patient characteristics that provide important additional predictive information to inform optimal treatment recommendations (eg, biomarkers and cognitive tasks) [12]. In addition, repetitive negative thinking, which emerged as a predictor of differential response, may be more validly assessed via methods other than conventional retrospective self-report questionnaires (eg, repeated, daily ecological momentary assessment [43,69]). Other relevant variables (eg, app use data, motivational variables, and involvement in other activities linked to better mental health) could be assessed in future studies. Second, our results emerged within a specific sample (school district employees), which did not have adequate representation of males, Black, Indigenous, and people of color, or individuals with low income. The sample is representative of Wisconsin in terms of race (83% of Wisconsinites are White) but includes a higher proportion of females. However, the gender difference in our sample is not surprising, given that females are more likely than males to (1) be employed as teachers [70] and (2) experience and seek treatment for depressive and anxiety symptoms [71]. Third, we were unable to conduct external validation by evaluating model prediction performance in an entirely new sample (eg, from another RCT). Fourth, as is common in digital therapies [48], a sizable subset of participants used the app for relatively few days. However, the results remained significant when restricting our analyses to subsets of participants who used the app for a longer period (Multimedia Appendix 1). Fifth, we did not include an active comparison condition. Our assessment-only control condition was not designed to control for placebo-related processes [72]. The methods demonstrated here may ultimately be most relevant in helping patients and clinicians decide between competing interventions that are intended to be therapeutic.

Conclusions and Future Directions

This study demonstrated the potential utility of data-driven approaches for informing personalized meditation app recommendations. A natural extension of this study is to conduct a prospective test of our algorithm using a doubly randomized design. For example, participants could be randomized to either (1) random treatment assignment (ie, treatment A or treatment B) or (2) be assigned to their algorithm-indicated treatment. To the extent that patient outcomes are significantly (and clinically meaningfully) better in the latter condition, the results would support the clinical benefits of algorithm-informed treatment recommendations (for a recent example of a similar design testing predictive matching of patients to therapists, refer to the study by Constantino et al [73]). In addition to comparing treatment packages, this design could be readily used to evaluate other customizable elements of HMP or other mHealth interventions. This may include assignment to receive various components or ordering of components within HMP, assignment to HMP or an alternative commonly used mHealth intervention (eg, CBT, behavioral activation, journaling, or mood tracking apps), or assignment to varying treatment intensities (eg, meditation practice frequency).

Other potentially fruitful future directions include evaluating a broader set of patient characteristics previously shown or hypothesized to predict the likelihood of responding to different interventions [5]. In addition, prediction models could be developed using data drawn from large naturalistic data sets evaluating mHealth interventions, as has been done for in-person psychotherapy and pharmacotherapy [65,74-76]. In addition to testing the utility of these models in “real-world” settings, naturalistic settings often provide large data sets relative to RCTs and thus can increase statistical power [28]. Ultimately, these approaches may gradually help supplement our reliance on trial and error for treatment selection with empirically supported, data-driven algorithms to objectively communicate expected benefits to individuals, allowing them to make well-informed decisions about which interventions are best for their needs.


This research was supported by National Center for Complementary and Integrative Health grant K23AT010879 (SBG), National Institute for Mental Health grant R01MH43454 (RJD), Chan Zuckerberg Initiative Grant 2020-218037 (RJD), a National Academy of Education/Spencer Postdoctoral Fellowship (MJH), and funding from the Wisconsin Center for Education Research (SBG). The first author (CAW) was partially supported by National Institute for Mental Health R01MH116969, National Center for Complementary and Integrative Health R01AT011002, the Tommy Fuss Fund, and a NARSAD Young Investigator Grant from the Brain & Behavior Research Foundation. SBG was partially supported by the Hope for Depression Research Foundation, Defeating Depression Award.

Conflicts of Interest

RJD is the founder, president, and serves on the board of directors for the nonprofit organization Healthy Minds Innovations, Inc. MJH has been a paid consultant at Healthy Minds Innovations, Inc for work unrelated to this study.

Multimedia Appendix 1


DOCX File , 440 KB

Multimedia Appendix 2

CONSORT e-HEALTH checklist.

PDF File (Adobe PDF File), 1068 KB

  1. Collins FS, Varmus H. A new initiative on precision medicine. N Engl J Med 2015 Feb 26;372(9):793-795 [FREE Full text] [CrossRef] [Medline]
  2. Paez J, Jänne PA, Lee J, Tracy S, Greulich H, Gabriel S, et al. EGFR mutations in lung cancer: correlation with clinical response to gefitinib therapy. Science 2004 Jun 04;304(5676):1497-1500. [CrossRef] [Medline]
  3. Rosell R, Carcereny E, Gervais R, Vergnenegre A, Massuti B, Felip E, Spanish Lung Cancer Group in collaboration with Groupe Français de Pneumo-CancérologieAssociazione Italiana Oncologia Toracica. Erlotinib versus standard chemotherapy as first-line treatment for European patients with advanced EGFR mutation-positive non-small-cell lung cancer (EURTAC): a multicentre, open-label, randomised phase 3 trial. Lancet Oncol 2012 Mar;13(3):239-246. [CrossRef] [Medline]
  4. Schwaederle M, Zhao M, Lee J, Eggermont A, Schilsky R, Mendelsohn J, et al. Impact of precision medicine in diverse cancers: a meta-analysis of phase II clinical trials. J Clin Oncol 2015 Nov 10;33(32):3817-3825. [CrossRef]
  5. Kessler RC, van Loo HM, Wardenaar KJ, Bossarte RM, Brenner LA, Ebert DD, et al. Using patient self-reports to study heterogeneity of treatment effects in major depressive disorder. Epidemiol Psychiatr Sci 2017 Feb;26(1):22-36 [FREE Full text] [CrossRef] [Medline]
  6. Norcross JC, Wampold BE. A new therapy for each patient: evidence-based relationships and responsiveness. J Clin Psychol 2018 Nov 18;74(11):1889-1906. [CrossRef] [Medline]
  7. Hastie T, Tibshirani R, Friedman J. The Elements Of Statistical Learning: Data Mining, Inference, And Prediction, Second Edition. Cham: Springer; 2017.
  8. Webb CA, Cohen ZD. Progress towards clinically informative data-driven decision support tools in psychotherapy. Lancet Digital Health 2021 Apr;3(4):e207-e208. [CrossRef]
  9. Boehmke B, Greenwell B. Hands-On Machine Learning with R. Boca Raton: Chapman and Hall/CRC; 2019.
  10. Dwyer DB, Falkai P, Koutsouleris N. Machine learning approaches for clinical psychology and psychiatry. Annu Rev Clin Psychol 2018 May 07;14(1):91-118. [CrossRef] [Medline]
  11. Aafjes-van Doorn K, Kamsteeg C, Bate J, Aafjes M. A scoping review of machine learning in psychotherapy research. Psychother Res 2021 Jan 29;31(1):92-116. [CrossRef] [Medline]
  12. Chekroud AM, Bondar J, Delgadillo J, Doherty G, Wasil A, Fokkema M, et al. The promise of machine learning in predicting treatment outcomes in psychiatry. World Psychiatry 2021 Jun 18;20(2):154-170 [FREE Full text] [CrossRef] [Medline]
  13. Cohen ZD, DeRubeis RJ. Treatment selection in depression. Annu Rev Clin Psychol 2018 May 07;14:209-236. [CrossRef] [Medline]
  14. Delgadillo J, Lutz W. A development pathway towards precision mental health care. JAMA Psychiatry 2020 Sep 01;77(9):889-890. [CrossRef] [Medline]
  15. DeRubeis RJ. The history, current status, and possible future of precision mental health. Behav Res Ther 2019 Dec;123:103506. [CrossRef] [Medline]
  16. DeRubeis RJ, Cohen ZD, Forand NR, Fournier JC, Gelfand LA, Lorenzo-Luaces L. The Personalized Advantage Index: translating research on prediction into individualized treatment recommendations. A demonstration. PLoS One 2014 Jan 8;9(1):e83875 [FREE Full text] [CrossRef] [Medline]
  17. Huibers MJ, Cohen ZD, Lemmens LH, Arntz A, Peeters FP, Cuijpers P, et al. Predicting optimal outcomes in cognitive therapy or interpersonal psychotherapy for depressed individuals using the personalized advantage index approach. PLoS One 2015 Nov 10;10(11):e0140771 [FREE Full text] [CrossRef] [Medline]
  18. Cohen ZD, Kim TT, Van HL, Dekker JJ, Driessen E. A demonstration of a multi-method variable selection approach for treatment selection: recommending cognitive-behavioral versus psychodynamic therapy for mild to moderate adult depression. Psychother Res 2020 Feb 11;30(2):137-150. [CrossRef] [Medline]
  19. Webb CA, Trivedi MH, Cohen ZD, Dillon DG, Fournier JC, Goer F, et al. Personalized prediction of antidepressant v. placebo response: evidence from the EMBARC study. Psychol Med 2018 Jul 2;49(07):1118-1127. [CrossRef]
  20. Jorm AF, Patten SB, Brugha TS, Mojtabai R. Has increased provision of treatment reduced the prevalence of common mental disorders? Review of the evidence from four countries. World Psychiatry 2017 Feb 26;16(1):90-99 [FREE Full text] [CrossRef] [Medline]
  21. Kohn R, Saxena S, Levav I, Saraceno B. The treatment gap in mental health care. Bull World Health Organ 2004 Nov;82(11):858-866 [FREE Full text] [Medline]
  22. Andersson G, Cuijpers P. Internet-based and other computerized psychological treatments for adult depression: a meta-analysis. Cogn Behav Ther 2009;38(4):196-205. [CrossRef] [Medline]
  23. Linardon J, Cuijpers P, Carlbring P, Messer M, Fuller-Tyszkiewicz M. The efficacy of app-supported smartphone interventions for mental health problems: a meta-analysis of randomized controlled trials. World Psychiatry 2019 Oct 09;18(3):325-336 [FREE Full text] [CrossRef] [Medline]
  24. Steinhubl SR, Muse ED, Topol EJ. Can mobile health technologies transform health care? JAMA 2013 Dec 11;310(22):2395-2396. [CrossRef] [Medline]
  25. Torous J, Andersson G, Bertagnoli A, Christensen H, Cuijpers P, Firth J, et al. Towards a consensus around standards for smartphone apps and digital mental health. World Psychiatry 2019 Feb;18(1):97-98 [FREE Full text] [CrossRef] [Medline]
  26. Chien I, Enrique A, Palacios J, Regan T, Keegan D, Carter D, et al. A machine learning approach to understanding patterns of engagement with internet-delivered mental health interventions. JAMA Netw Open 2020 Jul 01;3(7):e2010791 [FREE Full text] [CrossRef] [Medline]
  27. Eysenbach G. The law of attrition. J Med Internet Res 2005 Mar 31;7(1):e11 [FREE Full text] [CrossRef] [Medline]
  28. Luedtke A, Sadikova E, Kessler RC. Sample size requirements for multivariate models to predict between-patient differences in best treatments of major depressive disorder. Clin Psychol Sci 2019 Jan 18;7(3):445-461. [CrossRef]
  29. Lagan S, D'Mello R, Vaidyam A, Bilden R, Torous J. Assessing mental health apps marketplaces with objective metrics from 29,190 data points from 278 apps. Acta Psychiatr Scand 2021 Aug 29;144(2):201-210. [CrossRef] [Medline]
  30. Wasil AR, Gillespie S, Shingleton R, Wilks CR, Weisz JR. Examining the reach of smartphone apps for depression and anxiety. Am J Psychiatry 2020 May 01;177(5):464-465. [CrossRef] [Medline]
  31. Hirshberg M, Frye C, Dahl C, Riordan K, Vack N, Sachs J, et al. A randomized controlled trial of a smartphone-based well-being training in public school system employees during the COVID-19 pandemic. J Educ Psychol 2022 Mar 17 [FREE Full text] [CrossRef]
  32. Ettman CK, Abdalla SM, Cohen GH, Sampson L, Vivier PM, Galea S. Prevalence of depression symptoms in US adults before and during the COVID-19 pandemic. JAMA Netw Open 2020 Sep 01;3(9):e2019686 [FREE Full text] [CrossRef] [Medline]
  33. Markowitz A, Bassok D, Bellows L, Doromal J, Michie M, Smith A. Understanding early educators’ responses to the coronavirus pandemic: lessons from large-scale surveys in Louisiana. In: Proceedings of the Association for Public Policy Analysis & Management Annual Meeting. 2020 Presented at: Association for Public Policy Analysis & Management Annual Meeting; Jul 23-24, 2020; Health Sciences Building.
  34. Markowitz A, Ehrlich S, Gassman-Pines A. Panel - child care in the time of covid-19: new evidence from child care business owners, early educators, and parents to inform recovery efforts (family and child policy). In: Proceedings of the APPAM Virtual Fall Research Conference. 2020 Presented at: APPAM Virtual Fall Research Conference; Nov 13, 2020; Washington, DC.
  35. Pilkonis PA, Choi SW, Reise SP, Stover AM, Riley WT, Cella D, PROMIS Cooperative Group. Item banks for measuring emotional distress from the Patient-Reported Outcomes Measurement Information System (PROMIS®): depression, anxiety, and anger. Assessment 2011 Sep;18(3):263-283 [FREE Full text] [CrossRef] [Medline]
  36. Dahl CJ, Wilson-Mendenhall CD, Davidson RJ. The plasticity of well-being: a training-based framework for the cultivation of human flourishing. Proc Natl Acad Sci U S A 2020 Dec 22;117(51):32197-32206 [FREE Full text] [CrossRef] [Medline]
  37. Goldberg SB, Imhoff-Smith T, Bolt DM, Wilson-Mendenhall CD, Dahl CJ, Davidson RJ, et al. Testing the efficacy of a multicomponent, self-guided, smartphone-based meditation app: three-armed randomized controlled trial. JMIR Ment Health 2020 Nov 27;7(11):e23825 [FREE Full text] [CrossRef] [Medline]
  38. Cohen S. Perceived stress in a probability sample of the United States. In: The Social Psychology of Health. Thousand Oaks, California: SAGE Publications; 1988.
  39. Roberti J, Harrington L, Storch E. Further psychometric support for the 10-item version of the perceived stress scale. J College Counsel 2006;9(2):135-147.
  40. Schalet BD, Pilkonis PA, Yu L, Dodds N, Johnston KL, Yount S, et al. Clinical validity of PROMIS depression, anxiety, and anger across diverse clinical samples. J Clin Epidemiol 2016 May;73:119-127 [FREE Full text] [CrossRef] [Medline]
  41. Collins GS, Reitsma JB, Altman DG, Moons KG. Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD). Ann Intern Med 2015 May 19;162(10):735-736. [CrossRef]
  42. Varoquaux G, Raamana PR, Engemann DA, Hoyos-Idrobo A, Schwartz Y, Thirion B. Assessing and tuning brain decoders: cross-validation, caveats, and guidelines. Neuroimage 2017 Jan 15;145(Pt B):166-179. [CrossRef] [Medline]
  43. Webb C, Swords C, Murray L, Hilt L. App-based mindfulness training for adolescent rumination: predictors of immediate and cumulative benefit. Mindfulness (N Y) 2021 Oct;12(10):2498-2509 [FREE Full text] [CrossRef] [Medline]
  44. Wetherill RR, Rao H, Hager N, Wang J, Franklin TR, Fan Y. Classifying and characterizing nicotine use disorder with high accuracy using machine learning and resting-state fMRI. Addict Biol 2019 Jul 27;24(4):811-821 [FREE Full text] [CrossRef] [Medline]
  45. Webb CA, Swords CM, Lawrence HR, Hilt LM. Which adolescents are well-suited to app-based mindfulness training? A randomized clinical trial and data-driven approach for personalized recommendations. J Consult Clin Psychol 2022 Sep;90(9):655-669 [FREE Full text] [CrossRef] [Medline]
  46. Watkins ER, Nolen-Hoeksema S. A habit-goal framework of depressive rumination. J Abnorm Psychol 2014 Feb;123(1):24-34. [CrossRef] [Medline]
  47. Watkins ER, Roberts H. Reflecting on rumination: consequences, causes, mechanisms and treatment of rumination. Behav Res Ther 2020 Apr;127:103573. [CrossRef] [Medline]
  48. Torous J, Nicholas J, Larsen ME, Firth J, Christensen H. Clinical review of user engagement with mental health smartphone apps: evidence, theory and improvements. Evid Based Ment Health 2018 Aug 05;21(3):116-119. [CrossRef] [Medline]
  49. Steyerberg E. Clinical Prediction Models: A Practical Approach to Development, Validation, and Updating Second Edition. Cham: Springer; 2019.
  50. Hirshberg MJ, Goldberg S, Davidson R. COVID and the Healthy Minds Program for Educators. Open Science Framework. 2020 Jun 10.   URL: [accessed 2022-10-24]
  51. R Core Team. R: a language and environment for statistical computing. R Foundation for Statistical Computing. Vienna, Austria: R Foundation for Statistical Computing; 2021.   URL: [accessed 2022-10-24]
  52. Webb CA. R code for "Personalized prediction of response to smartphone-delivered meditation training" submission. Open Science Framework. 2021 Jul 27.   URL: [accessed 2022-10-24]
  53. Ehring T, Zetsche U, Weidacker K, Wahl K, Schönfeld S, Ehlers A. The Perseverative Thinking Questionnaire (PTQ): validation of a content-independent measure of repetitive negative thinking. J Behav Ther Exp Psychiatry 2011 Jun;42(2):225-232 [FREE Full text] [CrossRef] [Medline]
  54. Topp CW, Østergaard SD, Søndergaard S, Bech P. The WHO-5 Well-Being Index: a systematic review of the literature. Psychother Psychosom 2015;84(3):167-176 [FREE Full text] [CrossRef] [Medline]
  55. Baer RA, Smith GT, Lykins E, Button D, Krietemeyer J, Sauer S, et al. Construct validity of the five facet mindfulness questionnaire in meditating and nonmeditating samples. Assessment 2008 Sep 29;15(3):329-342. [CrossRef] [Medline]
  56. Cyranowski JM, Zill N, Bode R, Butt Z, Kelly MA, Pilkonis PA, et al. Assessing social support, companionship, and distress: National Institute of Health (NIH) toolbox adult social relationship scales. Health Psychol 2013 Mar;32(3):293-301 [FREE Full text] [CrossRef] [Medline]
  57. Raes F, Pommier E, Neff KD, Van Gucht D. Construction and factorial validation of a short form of the self-compassion scale. Clin Psychol Psychother 2011 Jun 08;18(3):250-255. [CrossRef] [Medline]
  58. Forman EM, Herbert JD, Juarascio AS, Yeomans PD, Zebell JA, Goetter EM, et al. The Drexel defusion scale: a new measure of experiential distancing. J Contextual Behav Sci 2012 Dec;1(1-2):55-65. [CrossRef]
  59. Steger MF, Frazier P, Oishi S, Kaler M. The meaning in life questionnaire: assessing the presence of and search for meaning in life. J Counsel Psychol 2006 Jan;53(1):80-93. [CrossRef]
  60. Stekhoven DJ, Bühlmann P. MissForest--non-parametric missing value imputation for mixed-type data. Bioinformatics 2012 Jan 01;28(1):112-118. [CrossRef] [Medline]
  61. Schwartz B, Cohen Z, Rubel J, Zimmermann D, Wittmann W, Lutz W. Personalized treatment selection in routine care: integrating machine learning and statistical algorithms to recommend cognitive behavioral or psychodynamic therapy. Psychother Res 2021 Jan;31(1):33-51. [CrossRef] [Medline]
  62. Team R. R: A Language and Environment for Statistical Computing. Vienna, Austria: R Foundation for Statistical Computing; 2021.
  63. Canty A, Ripley B. boot: Bootstrap R (S-Plus) Functions. R package version. 2021.   URL: [accessed 2022-10-21]
  64. Hayes AF, Matthes J. Computational procedures for probing interactions in OLS and logistic regression: SPSS and SAS implementations. Behav Res Method 2009 Aug;41(3):924-936. [CrossRef]
  65. Webb CA, Cohen ZD, Beard C, Forgeard M, Peckham AD, Björgvinsson T. Personalized prognostic prediction of treatment outcome for depressed patients in a naturalistic psychiatric hospital setting: a comparison of machine learning approaches. J Consult Clin Psychol 2020 Jan;88(1):25-38 [FREE Full text] [CrossRef] [Medline]
  66. Buckman J, Cohen Z, O'Driscoll C, Fried E, Saunders R, Ambler G, et al. Predicting prognosis for adults with depression using individual symptom data: a comparison of modelling approaches. Psychol Med 2021 May 06:1-11. [CrossRef]
  67. Flett JA, Hayne H, Riordan BC, Thompson LM, Conner TS. Mobile mindfulness meditation: a randomised controlled trial of the effect of two popular apps on mental health. Mindfulness 2018 Oct 31;10(5):863-876. [CrossRef]
  68. Lipton R, Stewart W, Stone A, Láinez MJ, Sawyer J, Disability in Strategies of Care Study group. Stratified care vs step care strategies for migraine: the Disability in Strategies of Care (DISC) Study: a randomized trial. JAMA 2000;284(20):2599-2605. [CrossRef] [Medline]
  69. Webb CA, Israel ES, Belleau E, Appleman L, Forbes EE, Pizzagalli DA. Mind-wandering in adolescents predicts worse affect and is linked to aberrant default mode network-salience network connectivity. J Am Acad Child Adolesc Psychiatry 2021 Mar;60(3):377-387 [FREE Full text] [CrossRef] [Medline]
  70. Characteristics of public school teachers. National Center for Education Statistics.   URL: [accessed 2021-10-06]
  71. Eaton NR, Keyes KM, Krueger RF, Balsis S, Skodol AE, Markon KE, et al. An invariant dimensional liability model of gender differences in mental disorder prevalence: evidence from a national sample. J Abnorm Psychol 2012 Feb;121(1):282-288 [FREE Full text] [CrossRef] [Medline]
  72. Torous J, Firth J. The digital placebo effect: mobile mental health meets clinical psychiatry. Lancet Psychiatry 2016 Feb;3(2):100-102. [CrossRef]
  73. Constantino MJ, Boswell JF, Coyne AE, Swales TP, Kraus DR. Effect of matching therapists to patients vs assignment as usual on adult psychotherapy outcomes: a randomized clinical trial. JAMA Psychiatry 2021 Sep 01;78(9):960-969 [FREE Full text] [CrossRef] [Medline]
  74. Bone C, Simmonds-Buckley M, Thwaites R, Sandford D, Merzhvynska M, Rubel J, et al. Dynamic prediction of psychological treatment outcomes: development and validation of a prediction model using routinely collected symptom data. Lancet Digital Health 2021 Apr;3(4):e231-e240. [CrossRef]
  75. Delgadillo J, Appleby S, Booth S, Burnett G, Carey A, Edmeade L, et al. The Leeds risk index: field-test of a stratified psychological treatment selection algorithm. Psychother Psychosom 2020;89(3):189-190. [CrossRef] [Medline]
  76. Webb C, Forgeard M, Israel E, Lovell-Smith N, Beard C, Björgvinsson T. Personalized prescriptions of therapeutic skills from patient characteristics: an ecological momentary assessment approach. J Consult Clin Psychol 2022 Jan;90(1):51-60 [FREE Full text] [CrossRef] [Medline]

CBT: cognitive behavioral therapy
CONSORT: Consolidated Standards of Reporting Trials
CV: cross-validation
HMP: Healthy Minds Program
mHealth: mobile health
MLQ: Meaning in Life Questionnaire
PAI: Personalized Advantage Index
PROMIS: Patient-Reported Outcomes Information System
PTQ: Perseverative Thinking Questionnaire
RCT: randomized controlled trial
REDCap: Research Electronic Data Capture
RMSE: root mean square error

Edited by R Kukafka; submitted 31.07.22; peer-reviewed by Z Li, Y Wang, M Kapsetaki; comments to author 02.09.22; revised version received 03.09.22; accepted 26.09.22; published 08.11.22


©Christian A Webb, Matthew J Hirshberg, Richard J Davidson, Simon B Goldberg. Originally published in the Journal of Medical Internet Research (, 08.11.2022.

This is an open-access article distributed under the terms of the Creative Commons Attribution License (, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research, is properly cited. The complete bibliographic information, a link to the original publication on, as well as this copyright and license information must be included.