Designing Effective Movement Digital Biomarkers for Unobtrusive Emotional State Mobile Monitoring

Abstract

Mobile sensing technologies and machine learning techniques have been successfully exploited to build effective systems for mental health monitoring and intervention. Various approaches have recently been proposed to effectively exploit contextual information such as mobility, communication and mobile usage patterns for quantifying users' emotional states and wellbeing. In particular, it has been shown that location information collected by means of smartphones can be successfully used to monitor and predict depression levels, as measured by means of standard scores such as PHQ-8.In this paper, we investigate the design of novel digital biomarkers based on the fine-grained characterization of the mobility patterns of a user, also considering the temporal dimension of their movements (e.g., sequence of places visited by them). We show that the proposed biomarkers have a statistically significant association with emotional states. We also demonstrate that emotional states have a stronger relationship with mobility patterns of weekdays compared to all days of a week. Finally, we discuss the challenges in using these biomarkers in the implementation of "emotion-aware" systems for digital health.

Investigating the Relationships between Mobility Behaviours and Indicators of Subjective Well–Being Using Smartphone–Based Experience Sampling and GPS Tracking

“…Hence, future research should aim to enrich mobility features with insights into corresponding activities assessed via other types of smartphone data or self–reports for naturalistic situation assessment (Harari et al, 2018), which would allow for a more nuanced understanding of how and when mobility behaviours are related to greater subjective well–being. For example, such research could capture people's subjective experience of the place they are in using the DIAMONDS framework of situational characteristics, which captures the psychology of a situation along eight distinct dimensions (Rauthmann et al, 2014).…”

“…where people spend their time; e.g. Mehl, Gosling, & Pennebaker, 2006; Rauthmann et al, 2014; Sandstrom et al, 2017). A likely explanation for this apparent gap in the literature is that mobility behaviours have been traditionally difficult to study in the context of daily life.…”

Evaluating Machine Learning Algorithms for Prediction of the Adverse Valence Index Based on the Photographic Affect Meter

“…Taking the day as the unit of time, features of dierent kinds are aggregated for the task of AV index prediction. Based on the recent literature [4,13,14], we add 7 GPS-based metrics: 1) total distance covered, 2) maximum 2-point separation, 3) number of dierent places visited by per-user tiled area grid approximation, 4) dierence in sequence of tiles covered, compared to previous day, 5) distance entropy, 6) number of non-routine clusters visited, 7) time spent on non-routine tiles. A selection has been made from the features proposed in previous studies in order to retain the metrics that least correlate with one another.…”

Towards Deep Learning Models for Psychological State Prediction using Smartphone Data: Challenges and Opportunities

“…We construct a neural network model and train it with the aim of predicting these stress reports based on predictor variables calculated for the same day. Eight spatio-temporal metrics based on the recent literature [13] are used as predictors: total distance covered in a day, maximum 2-point displacement in a day, distance standard deviation, number of different areas visited by tiles approximation, total spatial coverage by convex hull, difference in sequence of tiles covered compared to previous day, difference in sequence of clusters visited compared to previous day, distance entropy. After calculating the metrics based on the GPS data, they are standardised on per-user basis.…”

“…To evaluate our model, we use the StudentLife dataset [22], which contains rich and multi-dimensional data collected from students at Dartmouth College over the course of a term. We consider eight GPS metrics, based on the recent literature [13], and four temporal variables to indicate weekends and the start, middle and end of the term. We evaluate the model using the cross-validation approach and present results comparing the importance of these two feature types.…”