JMIR Publications

ABSTRACT

Background:

Sensors are increasingly used to capture unobtrusively and continuously free-living physical activity from participants to evaluate health education interventions and programs by objectively analysing physical activity behaviour changes. The rich granularity of sensor data offers a great potential for analysing patterns and changes in physical activity behaviours. Studies have been making greater use of specialised machine learning and data mining techniques to detect, extract and analyse these patterns in recent years, helping to better understand how participants’ physical activity evolves.

Objective:

This systematic review aims to identify and synthesise the various data mining techniques used to analyse changes in physical activity behaviours from sensors data in health education and health promotion intervention studies. We address two main research questions: (1) What are the current data mining techniques used for PA sensor data to detect behaviour changes in health education or health promotion contexts? (2) What are the challenges and opportunities of mining PA sensor data for detecting physical activity behaviour changes?

Methods:

A systematic review was conducted in May 2021 using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. We queried ACM, IEEE Xplore, ProQuest, Scopus, Web of Science, ERIC and Springer literature databases for peer-reviewed references related to wearable machine learning to detect physical activity changes in health education. 4438 references were initially retrieved from the databases. After removing duplicates and screening titles and abstracts, 285 references were full-text reviewed, resulting in 19 articles.

Results:

The main sensors used were accelerometers collecting data between 4 days to 1 year (median 10 weeks) and the number of participants varying between 10 and 11615 (median 74). Data preprocessing was mainly carried out by proprietary software, generally resulting in steps count and time spent in physical activity aggregated predominantly at a daily or minute level. The main features that serve as input for data mining models were mostly descriptive statistics of the preprocessed data. Preferred data mining methods were classifiers, clusters, and decision-making algorithms. Finally, the resulting data mining models focused on personalisation, support for self-reflection, and analysis of PA behaviours.

Conclusions:

In the last ten years, an increasing number of mining techniques for sensor data have been employed to detect behaviour changes in health education and promotion contexts. These generally require the generation of new features and detection methods. Where the population sample size and the recording time is sufficient, this may gain general and long-term insights into the evolution of the population’s behavioural changes. It also appears that adding complementary information to the physical activity data can help create more accurate models and provide better interpretation of these, allowing to generate more precise personalised feedback and support for participants. Exploring different data aggregation levels could help detect more subtle and long-term behavioural changes. It is also necessary to make transparent, explicit and standardised the processing procedures to establish best practices and make the detection methods easier to understand, scrutinise and reproduce.