Advancing Precision Mental Health Across Cultures: A Multimodal Ecological Momentary Assessment Study

Improving the Prediction of Early Treatment Responses, Dropout, and Outcome using Daily-Life Data

Psychotherapeutic approaches have repeatedly been shown to be effective, with effects comparable to those of pharmacological treatment and, in some cases, showing greater durability over time (e.g., Cuijpers et al., 2023). At the same time, overall treatment success remains modest, with more than half of patients not responding sufficiently to psychological treatment and many discontinuing therapy prematurely (Cuijpers et al., 2021; Lutz et al., 2021). Moreover, access to psychotherapeutic care remains particularly limited in low- and lower-middle-income countries, where only one in 27 people with a depressive disorder receive adequate treatment (Thornicroft et al., 2017). This highlights the continued need to better understand, across diverse settings, which patients are likely to benefit from specific interventions, which patients may be at risk of insufficient response, and how treatment can be adapted more effectively when progress is limited (Lutz et al., 2022; Nye et al., 2023).

To date, research on the prediction of treatment trajectories and outcome has relied predominantly on retrospective questionnaire data collected before or repeatedly during treatment. These approaches have provided important insights and shown that treatment trajectories can be predicted from pre-treatment information or session-by-session reports (e.g., Constantino et al., 2021; Lees & Delgadillo, 2026). However, retrospective self-reports are limited by recall bias and their reduced capacity to capture experiences in real time and natural contexts (e.g., Myin-Germeys et al., 2018). From a precision mental health perspective, which aims to advance personalized psychological care, improving the temporal accuracy of assessments may strengthen prediction models and, in turn, support more tailored treatment decisions (e.g., Lutz et al., 2025). 

How Ecological Momentary Assessment Can Complement Retrospective Measures

One approach to improve the precision of measurements is Ecological Momentary Assessment (EMA), also referred to as Experience Sampling Method (ESM; e.g., Hektner, 2007) or Ambulatory Assessment (AA; e.g., Trull & Ebner-Priemer, 2013). EMA involves the repeated assessment of affective, cognitive, or behavioral processes in individuals’ natural environments (Stone & Shiffman, 1994). The term EMA encompasses different assessment methods, including active self-reports and passive sensing of physiological indicators, such as heart rate (Shiffman et al., 2008). By capturing experiences in real-world contexts and close to the moment of occurrence, EMA allows researchers to capture fluctuations and short-term dynamics in daily life, thereby reducing retrospective distortions and increasing the ecological validity of the collected information (e.g., Ben-Zeev et al., 2012; Wright & Woods, 2020). 

Over the past decade, EMA-based research has expanded substantially across clinical psychology and psychotherapy research (Fritz et al., 2024), particularly regarding the prediction of symptom and treatment trajectories as well as treatment outcome (see Mink et al., 2025 for a systematic overview). Initial studies suggest that self-reported EMA data can improve the prediction of early treatment responses (e.g., Husen et al., 2016), treatment outcome (e.g., Hehlmann et al., 2025) and dropout (e.g., Lutz et al., 2018) beyond initial impairment assessed at baseline. Moreover, recent studies implementing passive sensing parameters yielded similar results when predicting treatment outcome (e.g., Hehlmann et al., 2021) and the severity of psychopathological symptoms (e.g., Jacobson et al., 2019; Siddi et al., 2023). However, it remains unclear to what extent the combination of EMA data modalities, including both active self-reports and passive sensing, can enhance the prediction of treatment trajectories and outcome beyond comprehensive baseline measures.

Integrating Context-Sensitive Perspectives Into Outcome Prediction

Although considerable research has examined how prediction of treatment courses and outcomes in psychotherapy can be improved, most of this work has been conducted in relatively homogeneous settings, limiting the generalizability of these models across sociocultural and socioeconomic contexts (Gómez Penedo et al., 2024). This limitation is particularly relevant given evidence that mental health problems are more prevalent in economically disadvantaged areas and that individuals from these contexts tend to show lower response rates to psychotherapy (Delgadillo et al., 2016). In addition, psychotherapy may be conceptualized and practiced differently across cultural contexts (Aafjes-van Doorn et al., 2020), while cultural adaptations of psychotherapeutic interventions have been associated with improved outcomes, underlining the importance of contextual sensitivity in both treatment and research (Anik et al., 2021; Lee et al., 2021). 

From this perspective, improving prediction in psychotherapy also requires examining whether predictive models remain valid across diverse contexts, including different cultural backgrounds, socioeconomic conditions, and healthcare settings. A model that performs well in one setting may not necessarily generalize to another. Integrating context-sensitive perspectives into prediction research may therefore help identify predictors that are robust across settings. 

The Present Project

In the context of these research findings, the introduced EMA study examines whether multimodal daily-life data can improve the prediction of treatment responses and overall outcome in psychotherapy. More specifically, the project investigates which EMA parameters assessed via self-reports and passive sensing at treatment onset can improve the prediction of early treatment responses, dropout, and treatment outcome beyond information collected through cross-sectional baseline assessments. In addition, the study examines whether these models generalize across culturally diverse treatment settings. This may help clarify not only whether daily-life data improve prediction, but also whether such models can support personalized treatment planning across cultural environments. The preregistration of the study can be accessed here: https://osf.io/ezpc3.

The study draws on data from 300 patients across three outpatient clinics in Trier, Germany, Buenos Aires, Argentina, and Osnabrück, Germany. At the beginning of treatment, patients report on their emotional and interpersonal experiences in daily life six times per day over a two-week period. At the same time, wearable devices (i.e., Garmin Vivosmart 4) are used to passively capture physiological and behavioral indicators such as heart rate, activity, and sleep. These repeated daily-life assessments are combined with baseline clinical information (e.g., initial impairment, treatment expectation, socioeconomic status) to examine whether multimodal EMA data improve the prediction of early treatment responses, dropout probabilities, and outcomes beyond cross-sectional baseline measures, and whether predictive performance remains comparable across the three study sites.

The aim of this work is not only to improve statistical prediction, but also to contribute to a more clinically meaningful understanding of treatment development. If EMA data allow the early identification of patients who are at risk of poor response or premature dropout, they may support more timely treatment adjustments and more targeted clinical decision-making. Moreover, if the cross-cultural validity of the algorithms is supported, much of the predictive knowledge generated in Europe and North America could be extended to other countries and cultural contexts, such as Latin America. This could have a substantial impact in settings with fewer resources for personalized mental health care. 

Current Status and Next Steps

Currently, the prediction models are being trained on the first half of the sample. Once data collection is complete, these models will be evaluated in an independent holdout sample, providing an out-of-sample evaluation of predictive accuracy and cross-cultural validity.

Conclusion

By integrating repeated self-reports and wearable-based indicators collected in patients’ natural environments, this study seeks to examine whether multimodal EMA data can enhance the prediction of early treatment response, dropout, and overall outcome beyond cross-sectionally assessed baseline measures. At the same time, its cross-cultural design addresses the important question of whether the predictive performance of such models is comparable across contexts and cultures. Ultimately, this work aims to contribute to early risk detection of poor treatment responses or premature dropout, more targeted treatment planning, and a more context-sensitive understanding of therapeutic change in culturally diverse settings.