Fig. 1: The diagram illustrating the analysis pipeline.

a The initial dataset consists of behavioural data from 215 participants across 27 mobile sensing categories in time intervals; an interval is defined as the period between two consecutive EMAs. The end-of-interval EMAs were taken as the ground truth mental health measures for the interval. All sensor data were originally in the format of three periods of the day (12 am–9 am, 9 am–6 pm, 6 pm–12 am) and were normalised (e.g., hour and second units transformed into minutes). b NMF was applied separately to each sensor data type for each interval, decomposing the data into a set of routine behaviours and their corresponding intensities over the interval. The number of routines was determined through leave-one-out cross-validation (LOOCV), yielding the lowest reconstruction error. The NMF routines retain the same data units as the original data. Routine variability was computed as the standard deviation of routine intensity across the interval. c For each sensing category, routine behaviours from all 215 participants were concatenated into a matrix to identify similar patterns for labelling. Specifically, the 25%, 50%, and 75% quantiles of each triplet item (corresponding to each period of the day) were computed, and labelling was based on the unique combinations of these results. This means a given sensing category could have a maximum of 27 routine types (3³). d An example of labelled routines for Number of Conversations (convo_num). Routine #1 means the participant had zero conversations in the morning (before 9a m), fewer than 50 (but more than zero) during the day (9 am–6 pm), and fewer than 25 in the evening (after 6 pm). e New matrices were obtained for all 215 participants’ labelled NMF routines for each interval. f For each participant, a generalised linear model (GLM) for anxiety and another for depression were built to associate routines and their variability with PHQ-2 and GAD-2 scores, respectively. g A large language model (OpenAI’s GPT-4o) interpreted an individual’s GLM summary to generate practical insights. h Subsequently, individuals were grouped if their GLMs shared the same set of themed routine categories, identified by averaging significant GLM coefficients across broader life aspects. Certain parts of this figure come from sources include Wikimedia Commons (the logo of OpenAI), The Noun Project (https://thenounproject.com) (icons indicating participants and sensing categories) under the Creative Commons Attribution 3.0 license.