Personalised modelling of routine variability and affective states.

Multimodal smartphone sensor data provide rich insights into real-world behavioural patterns associated with anxiety and depression symptoms. We propose that variability in daily routines across life aspects may act as a personalised digital marker, linking behaviours to self-reported affective states. Using non-negative matrix factorisation (NMF), we decomposed mobile sensing data into individual-specific routines and their weekly variability. Generalised linear models (GLMs) were built per individual to associate variability from specific sensing categories with anxiety or depression states. Using a large language model (LLM; GPT-4o), we translated the modelling results into more accessible language that potentially helps individuals engage with their own routine phenotypes for self-regulation insights. Population-level grouping of GLMs revealed significant between-group differences in mental health measures. The LLM output shows a potential to uncover personalised behavioural patterns and support self-understanding of mood-related drivers. This approach may also inform group-based interventions by identifying individuals likely to benefit from shared treatment strategies based on routine similarity.