Physical inactivity and obesity contribute signi cantly to several health problems and have been a topic of research amongst the public health research community. Objective measurements of physical behaviour using body-worn sensors have radically uplifted the method of collecting physical behaviour data from a large population and opened the possibility of analysing these objective recordings using machine learning methods to gain useful insights such as identifying di erent physical activities and the duration of time spent in each activity, among others. We aim to address the theme of physical activity from both the public health and computer science perspective by rst determining the best strategy to increase the accuracy of human activity classi cation from body-worn sensor data collected in cohort studies using the same data collection protocol as HUNT41 and strategically using objectively measured physical activity data to identify clusters of di erent physical behaviour phenotypes such that intra-cluster similarity is maximized. Identifying physical behaviour phenotypes can allow us in the future to design tailored interventions for users to self-manage their daily physical activity routines. Additionally, we consider the compositional nature of the physical behaviour data we have and plan to employ suitable methodologies which allow us to model the compositional co-dependency between physical behaviours and sleep.
Maintaining an active lifestyle has been shown to be essential in leading a healthy
life and has several positive e ects such as improved mental health [
In addition to regular physical activity, sleep duration and sleep quality are
important but highly volatile factors for sound health and have a strong
compositional co-dependency with daily physical behaviour. Although there is no
doubt that physical behaviour and sleep have a profound impact on health, it
has recently become evident that our current understanding of the health e ect
of these entities is awed by serious methodological shortcomings. In particular,
this pertains to the use of self-reported data, which is prone to bias and
misclassi cation [
This research builds on the requirement to generate classi cation models which are general enough to provide accurate classi cation of distinct physical behaviours from sensor data collected in cohort studies with the same data collection protocol as HUNT4 and furthermore, to e ectively utilise the objective data to identify di erent physical behaviour phenotypes. HUNT4 is the fourth round of the HUNT2 cohort study which collects large amount of both subjective as well as objective health data and objective physical activity data collected through body-worn accelerometers (which forms the target dataset in this research). 2
The overall goal is to achieve a high classi cation accuracy in classifying basic human activities from sensor data streams collected during HUNT4 and utilise this information to correctly estimate the time spent by an individual in each of the six di erent lower level physical activities: lying, sitting, standing, walking, running, cycling. Furthermore, these objective duration of physical activities of the participants are intended to be utilised to identify phenotypes of physical activities, which is important in order to design personalised and sustainable digital interventions for users for self-managing daily physical activity routines. 2 https://www.ntnu.no/hunt/ 2.1
Foreseen challenges { How to improve performance of machine learning classi ers for identifying human behaviour accurately using sensor data collected during HUNT4? { How to identify sleep duration from multiple sensor data streams using machine learning methods? { How to analyse the quality of clusters (physical behaviour phenotypes) obtained using di erent clustering methods? { How to design tailored interventions of daily physical activity routines for users using existing physical behaviour pro les?
To address the rst challenge, we will use di erent training datasets, including one wherein the data has been collected in out-of-lab settings and evaluate their performance by comparing with other model(s) trained using dataset(s) collected in in-lab settings. To address the second challenge, we will implement methodologies to identify sleep patterns and duration from Polysomnographic (sleep) recordings. Finally we plan to implement a similarity-based method for clustering the physical activity pro les of the HUNT4 participants. To analyse the quality of clusters obtained, we plan to implement state-of-the-art clustering methods and compare the performance of the similarity-based method with them. Furthermore, we plan to utilise clustering evaluation measures for comparing the cluster quality of the implemented methods to generate a quality assessment. The resulting clusters are expected to localise our search for the physical behaviour phenotypes in our dataset. 2.2
Overall, the main aim of this thesis is the exploration of machine learning methods for the HUNT4 physical behaviour raw data recordings. We expect to build a toolbox that allows: { Analysis of the body-worn accelerometer data using machine learning methods. { Analysis of the high-resolution physical behaviour data using CBR and Compositional data analysis. 3 3.1
A thorough review of the existing state-of-the-art and literature of human activity recognition systems forms the basis of this research. Identifying gaps in the existing work enables us to make improvements on previously untouched topics. Additionally, an overview of the literature and methodologies for Compositional Data Analysis is required for understanding the co-dependency between various physical behaviours and their in uence on one another since the data we are using is compositional in nature. 3.2
Several state-of-the-art methods such as Random Forest, k-NN, SVM and
Hidden Markov Models among others have been successfully implemented for
human activity recognition systems and have been shown to be quite e ective in
identifying di erent physical behaviour. Neural Network implementations such
as Recurrent Neural Networks [
For gaining insights from the public health dataset we have, we plan to utilise di erent methodologies including CBR and Compositional Data analysis. CBR has been used in our work for developing the similarity measures for assessing the similarity amongst and nding similar physical behaviour pro les of the participants of HUNT4. Furthermore, we will utilize the knowledge intensive similarity used in CBR to generate clusters of semantically similar behaviour pro les in order to localise our search for the physical behaviour phenotypes. Additionally, we will employ methodologies for compositional data analysis to estimate the e ect of increase or decrease in sleep and/or sedentary time on the composition of daily activities since an increase in the duration of one behaviour, say sitting, will necessarily reduce duration of at least one other behaviour because the total time in any given day is xed at 24 hours. Thus, whether a change in one type of behaviour is bene cial or harmful for health depends on the compensatory shifts in other behaviours. This data analysis methodology will be studied extensively along with various other related research articles to get a better understanding of it and how it can be applied to gain insights into the public health data and thereby utilize it to design tailored and sustainable digital interventions for the users to self-manage their daily physical activity routines . 4
We have presented our methodology for developing the similarity measures in
myCBR for assessing the similarity amongst the physical behaviour pro les at
ICCBR 2018 [