Personalized mood prediction from patterns of behavior collected with smartphones

Brunilda Balliu, Chris Douglas, Darsol Seok, Liat Shenhav, Yue Wu, Doxa Chatzopoulou, William Kaiser, Victor Chen, Jennifer Kim, Sandeep Deverasetty, Inna Arnaudova, Robert Gibbons, Eliza Congdon, Michelle G. Craske, Nelson B. Freimer, Eran Halperin, Sriram Sankararaman, Jonathan FlintNpj Digit. Medicine 2024 · DOI

TL;DR

This paper presents a novel method for predicting future mood in individuals with depression using passively collected smartphone data. By combining advanced statistical modeling with high-quality longitudinal mood assessments, they achieve significant improvements in prediction accuracy, offering a promising step towards real-time, personalized mental health monitoring.

What Problem Does This Paper Address?

Major Depressive Disorder (MDD) is a leading cause of disability, yet it's often underdiagnosed and difficult to monitor effectively. Traditional clinical assessments rely on subjective interviews with limited inter-rater reliability and are not ideal for continuous monitoring. While digital phenotypes from smartphones offer a promising avenue for passive data collection, previous studies have struggled with replicability, detecting clinically relevant events, and using validated depression measures. This paper tackles these limitations by developing a robust, personalized system for predicting mood weeks in advance, aiming to bridge the gap between digital data and actionable clinical insights.

Methods

The core of this research lies in its sophisticated approach to modeling. The authors employed:

  • Computerized Adaptive Testing (CAT-DI): This method provides high-quality, validated longitudinal mood assessments by tailoring questions to the individual's responses in real-time. This is a significant improvement over traditional self-report questionnaires for longitudinal studies.
  • Idiographic Modeling: Instead of a one-size-fits-all approach, they developed individualized prediction models for each participant. This accounts for the significant inter-individual variability in how behavior relates to mood.
  • Cubic Spline Interpolation: To handle the sparse nature of mood assessments (typically weekly or bi-weekly), they used cubic splines to interpolate and infer a continuous "latent trait" of depression between actual measurements. This allows for a more granular understanding of mood fluctuations.
  • Digital Behavioral Phenotypes: Data from smartphone sensors (location, screen activity, call/text logs) were processed into various features (e.g., activity levels, social interaction, sleep quality, device usage). These features were then used to predict future mood.
  • Elastic Net Regression: This statistical technique was used for building the prediction models, balancing the strengths of Lasso and Ridge regression for feature selection and regularization.

Datasets

The study utilized data from 183 individuals experiencing mild to severe symptoms of depression. These individuals were part of the Screening and Treatment for Anxiety and Depression (STAND) study at UCLA.

  • Mood Data: Collected via the Computerized Adaptive Testing Depression Inventory (CAT-DI) over periods of up to 40 weeks.
  • Behavioral Data: Passively collected from smartphone sensors (via the AWARE framework) over periods of up to 40 weeks, resulting in approximately 29,254 days of behavioral assessment.

This dataset is notable for its longitudinal nature and the combination of validated mood assessments with rich, passively collected behavioral data.

Key Insights and Ideas

  • Individualized Prediction is Key: The most striking insight is the significant superiority of idiographic (individualized) models over nomothetic (population-level) models. This underscores that how behavior relates to mood is highly personal.
  • Passive Sensing is Powerful: Smartphone sensor data, when analyzed appropriately, can capture subtle behavioral changes that are strongly correlated with mood states.
  • Mood is a Latent Trait: The use of cubic splines to infer a continuous "latent trait" of depression between discrete assessments is a clever way to leverage sparse data and capture dynamic mood changes.
  • Beyond Simple Correlations: The study moves beyond simple correlations to build predictive models, demonstrating the practical utility of digital phenotypes for forecasting mood.

Key Results

  • High Prediction Accuracy: The idiographic models achieved a prediction accuracy of depression severity up to three weeks in advance with R² ≥ 80%.
  • Significant Error Reduction: The personalized models resulted in a 65.7% reduction in prediction error compared to a baseline model that only used past depression severity.
  • Individual-Level Success: The idiographic approach successfully predicted future mood for 79.0% of individuals with significant accuracy (R>0 and FDR < 5%).
  • Feature Importance Varies: While "study day" was a common predictor, specific behavioral features (e.g., variation in social interaction, sleep patterns, time spent at home) were highly predictive but varied significantly in their importance across individuals.
  • Feasibility Demonstrated: The study successfully demonstrated the feasibility of obtaining high-quality longitudinal mood assessments and predicting symptom severity weeks in advance using passively collected digital behavioral data.

Limitations and Future Work

  • Idiographic Model Limitations: Fitting models separately for each individual might limit statistical power. The assumption of a linear relationship between behavior and mood might not hold for all individuals.
  • Response Bias: Repeated use of CAT-DI over many months could potentially lead to limited response bias.
  • Data Missingness: While imputation methods were used, the assumption of missing at random (MAR) is hard to definitively test and might be violated if participants with less engagement received fewer reminders.
  • Generalizability: The study population (UCLA students) might limit the generalizability of findings to broader demographics.
  • Future Directions: The authors suggest integrating data from wearable devices and electronic health records to further enhance prediction accuracy. They also envision this technology evolving into a real-time screening and detection system for clinicians.

Summary updated

BibTeX
@article{balliu2024personalized,
  author = {Brunilda Balliu and Chris Douglas and Darsol Seok and Liat Shenhav and Yue Wu and Doxa Chatzopoulou and William Kaiser and Victor Chen and Jennifer Kim and Sandeep Deverasetty and Inna Arnaudova and Robert Gibbons and Eliza Congdon and Michelle G. Craske and Nelson B. Freimer and Eran Halperin and Sriram Sankararaman and Jonathan Flint},
  title = {{Personalized mood prediction from patterns of behavior collected with smartphones}},
  journal = {Npj Digit. Medicine},
  year = {2024},
  doi = {10.1038/s41746-024-01035-6},
}
Figure