Frontiers in Public Health (Oct 2015)
From Mobile Phone Monitoring of Depressive States using GPS Traces Analysis to Data-Driven Behaviour Change Interventions
Abstract
Background Depressive disorders do not only affect the personal life of individuals and their social circles but it has also a strongly negative economic impact. For example, according to a recent study by the European Depression Association, workers in the United Kingdom suffer higher levels of depression than those anywhere else in Europe. The survey found that 1 in 10 employees had taken time off at some point in their working lives because of depression problems. We believe that the support provided by new mobile technologies can help to tackle this problem providing new ways for supporting both patients and healthcare officers, possibly through the automatic delivery of behaviour interventions. Aims Existing interview-based studies in the literature have shown that depression leads to a reduction of mobility and activity levels. The goal of our project is to investigate how mobile phones can be used to collect and analyse mobility patterns of individuals in order to quantitatively understand how mental health problems affect their daily routines and behaviour and how potential changes can be automatically detected. More specifically, we investigate the design, implementation and evaluation of analytical techniques for studying the correlation between patterns of human mobility and emotional states. More specifically, the aims of our project is to provide answers to the following questions: a) is there any correlation between mobility patterns extracted from GPS traces and depressive mood?; b) is it possible to devise unobtrusive smartphone applications that collect and exploit only mobility data in order to automatically infer a potential depressed mood of the user over time?; c) if this is possible, can we devise behaviour interventions based on the inferences we can derive from the mobile phones?; d) what are the ethical and practical implications related to the automatic delivery of behaviour intervention in the case of depressive mood disorders without the involvement of mental health professionals? Methods During our project, we designed an energy-efficient Android application to collect mobility data and assess the presence of depressed mood disorders by analysing mobility traces. We deployed the application and we collected data from 28 users. We also introduced a set of mobility metrics that can be extracted from the mobility traces of the users and, using the ground truth data collected by means of the Android application. More specifically, we considered the following mobility traces: We extracted the following mobility metrics: 1) The total distance covered; 2) The maximum distance between two locations; 3) The radius of gyration; 4) The standard deviation of the displacements; 5) The maximum distance from home; 6) The number of different places visited; 7) The number of different significant places visited; 8) The routine index. We trained and evaluated personalised and general machine learning models to predict PHQ score changes from mobility metrics variations. Results During our experiments, we identified a significant correlation between the changes of such metrics and the variations in the PHQ score. The correlation ranges from 0.336 (p-value: 0.181) to 0.432 (p-value: 0.069) when the mobility metrics are computed over a period of 14 days. With respect to the personalised prediction models, we obtained very good prediction accuracies. For example, when the mobility metrics are computed over a period of 14 days, the general model achieves sensitivity and specificity values of 0.74 and 0.78, respectively, whereas the average sensitivity and specificity values of the personalized models are 0.71 and 0.87, respectively. We are currently exploring how to exploit these findings. We are developing models for taking decisions based on the machine learning models we investigated. We are developing solutions for the automatic delivery of information based on the observed behaviour and inferred depressive states. In other words, we are designing feedback based systems, where the actual effectiveness of the intervention, measured by the change in the mood of the individual, is constantly monitored in order to take the next intervention decision. An open question is how to establish whether the observed mood change is actually the result of the intervention or external causes contribute to it. Conclusions We have demonstrated that it is possible to observe a significant correlation between mobility patterns and depressive mood using data collected by means of smartphones. We have also shown that it is possible to develop inference algorithms as a basis for unobtrusive monitoring and prediction of depressive mood disorders. The key open question is how to exploit the correlations between mobility metrics and depressive states we observe in the data. We are currently exploring a variety of possible solutions for enabling automatic delivery of behaviour intervention through real-time analysis of the sensed data. The focus of this initial work is on a specific modality, i.e., GPS location, but the results of this work can be indeed exploited to build more complex system based on the analysis of data extracted by means of other sensors, such as accelerometers, and other sources of information, such as call and SMS logs. We indeed plan to use the application in future studies that will focus on specific populations, such as clinically-diagnosed depressed individuals. Ethical considerations are also an important part of our investigation: we believe that the potential risks associated to the delivery of incorrect behaviour interventions should be analysed in depth. A possible solution might consist in mixed intervention methods, based on the automatic delivery of behaviour interventions by means of mobile phones with the involvement of mental healthcare officers and clinicians, at least in case of mild and severe depressive cases.
Keywords
