Activity recognition in mental health monitoring using multi-channel data collection and neural network

Abstract

Ecological momentary assessment (EMA) methods can be used to extract context related information by studying a subject’s behaviour in an environment in real-time. In mental health EMA can be used to assess patients with mental disorders by deriving contextual information from data and provide psychological interventions based on the behaviour of the person. With the advancements in technology smart devices such as mobile phone and smartwatch can be used to collect EMA data. Such a contextual information system is used in SyMptOMS, which uses accelerometer data from smartphone for activity recognition of the patient. Monitoring patients with mental disorders can be useful and psychological interventions can be provided in real time to control their behavior. In this research study, we aim to investigate the effect of multi-channel data on the accuracy of human activity recognition using neural network model by predicting activities based on data from smartphone and smartwatch accelerometer sensors. In addition to this the study investigates model performance for similar activities such as SITTING and LYING DOWN. Tri-axial accelerometer data is collected at the same time from smartphone and smartwatch using a data collection application. Features are extracted from the raw data and then used as input to a neural network. The model is trained for single data input from smartphone and smartwatch as well the data from sensor fusion. The performance of the model is evaluated by using test samples from collected data. Results show that model with multi-channel data achieves a higher accuracy of activity recognition than the model with only single-channel data source.