Customized Analysis of Daily Routine Variability and Emotional States

Overview

This study utilized multimodal smartphone sensor data to identify individual-specific daily routine variability linked to anxiety and depression symptoms. Personalized models revealed associations between routine fluctuations and affective states, while large language model translations enhanced interpretability and potential self-regulation.

Background

Understanding real-world behavioral manifestations of anxiety and depression remains challenging due to reliance on self-reported questionnaires prone to bias. Multimodal sensor data from smartphones offer granular, longitudinal behavioral insights that capture routine patterns across life aspects. Routines, both primary (e.g., sleep-wake cycles) and secondary (e.g., exercise, mealtimes), influence mental health, with disruptions linked to symptom severity. Digital phenotypes derived from such data enable personalized virtual representations, facilitating mechanistic insights into behavioral and psychological interactions underlying mental health.

Data Highlights

Using non-negative matrix factorisation (NMF), individual-specific routines and their weekly variability were extracted from smartphone sensor data. Generalised linear models (GLMs) per individual associated variability in specific sensing categories with anxiety and depression states. Population-level grouping of GLMs revealed significant between-group differences in mental health measures. A large language model (GPT-4o) translated these results into accessible language to support self-understanding and engagement.

Key Findings

• Variability in daily routines across multiple life aspects correlates with self-reported anxiety and depression symptoms.
• Personalized GLMs identified specific routine categories whose fluctuations associate with affective states on an individual basis.
• Population-level grouping based on routine variability patterns revealed statistically significant differences in mental health measures between groups.
• Large language model translations of modeling outputs improved interpretability and may facilitate user engagement and self-regulation.
• This approach supports identification of individuals likely to benefit from shared treatment strategies based on routine similarity.

Clinical Implications

Clinicians may leverage personalized digital phenotyping from smartphone sensor data to monitor routine variability as a marker of anxiety and depression. Translating complex modeling results into accessible language can empower patients to understand and regulate mood-related behavioral patterns. Grouping patients by routine similarity may inform tailored or group-based interventions to optimize mental health outcomes.

Conclusion

Personalized analysis of daily routine variability using multimodal sensor data provides valuable insights into anxiety and depression symptomatology. Coupling this with large language model interpretation enhances clinical utility and patient engagement, offering a promising avenue for personalized mental health monitoring and intervention.

References

1. Multimodal smartphone sensor data and mental health research
2. Non-negative matrix factorisation and GLMs for behavioral data analysis
3. Large language models for translating complex clinical data