Connecting physical activity with context and motivation: a user study to define variables to integrate into mobile health recommenders

Abstract

In this paper, we aim to improve existing health recommender systems by defining relevant contextual and motivational variables to recommend physical activities and collect appreciation feedback. Current health recommenders do not sufficiently include users’ context and motivational theory when personalizing health suggestions. To bridge these gaps, we conducted a 21-day longitudinal user study with 36 participants using our Android app with collected sensor data and Ecological Momentary Assessments to collect daily activities, mood, and motivation. This study resulted in a dataset of 724 activities. Two approaches to determine feature relevance were followed: variable importances analysis on 40 input variables, and statistical analysis of mean differences in outcome variables across contexts. Our findings suggest recommending activity duration, intensity, location, and type by incorporating: company, situation (e.g., free time or work), happiness, calmness, energy level, physical complaints, and motivation. As such, we propose opportunities for future health recommenders to integrate these data with contextual pre-filtering techniques, extended with our suggestions for automatically collected weather, location types, step count, and time. We also propose to use mood and motivation as appreciation feedback to focus on user well-being and boost motivation.

Appendices

Appendix A Selected questions from the behavioural regulation in exercise questionnaire (BREQ)

To keep the app user friendly, a selection of questions from the BREQ was chosen. However, some questions of the same BREQ construct are almost exactly the same (e.g., “I exercise because others say I should” and “I exercise because other people say I should” in external regulation (Cid et al. 2018)). Therefore, we chose two questions per construct, based on avoiding similar questions, and with their internal consistency checked using Cronbach’s alpha. This resulted in the following 12-item questionnaire:

• Amotivation: “I don’t see why I should have to exercise”, and “I think that exercising is a waste of time” (Markland and Tobin 2004)

• External regulation: “I exercise because others say I should”, and “I feel under pressure from others to exercise” (Wilson et al. 2006)

• Introjected regulation: “I feel guilty when I don’t exercise”, and “I feel a failure when I haven’t exercised” (Wilson et al. 2006)

• Identified regulation: “I value the benefits of exercise”, and “It’s important to me to exercise regularly” (Wilson et al. 2006)

• Integrated regulation: “I consider exercise to be a part of my identity”, and “I consider exercise a fundamental part of who I am” (Wilson et al. 2006)

• Intrinsic regulation: “I exercise because it’s fun”, and “I enjoy my exercise sessions” (Wilson et al. 2006)

Appendix B Data preprocessing criteria

Participants were informed about the study criteria at the beginning of the study. Based on these criteria, the following 25 participants were excluded during data preprocessing of the collected data:

• The pre- and post-test questionnaires should be filled in: 20 participants failed to fill in the post-test questionnaire

• 21 different days of one or more submits should be sent between the 1st of November, 2021 and the end of the first week of December, 2021: 1 participant submitted until the 11th of December which was too late

• Of those 21 days, there should be no more than two consecutive days with a submit without any PA (because every kind of PA counts, such as walking in the supermarket): 4 participants submitted data with the option “no activities performed today” more than two consecutive days

This preprocessing resulted in 36 participants who met our criteria. For these participants, their submitted data points were also preprocessed following these steps (with the amount of affected data points in parentheses):

• Remove “no activities performed today” data point if this was submitted on the same day when a PA record of that same user was also submitted (49)

• Remove data point if submitted less than 30 s after previous submit of that same user (3)

• Outliers for step count: if the majority of the detected step counts of a participant does not have a value, we assume a malfunctioning accelerometer and mark all step counts of that participant as missing (3 participants)

• Outliers for PA duration were determined based on the 99.5 percentile, resulting in three data points to be removed: one with a duration of 15 h, and two with a duration of 23 h (3 points in total)

• Remove data points with manually typed PA types involving only sedentary activities, since these are not classified as PA by World Health Organization (2020a): “reading” (5), “Zoom conversation” (4), and “date” (1)

Removal of these data points resulted in a total of 1427 valid submits. Since people could type their activities themselves, spelling errors were adjusted and remaining PA types were manually re-categorized:

• “Walking” for manually entered “went shopping” (2), “groceries” (3), “shopping” (1), and “going to a market” (1)

• “Padel” for manually entered “padellen” (1), and “paddelen” (1)

• “Fitness” for manually entered “Exercise in gym” (2), “gym” (8), and “eliptical” (2)

• “Weight training” for manually entered “Lifting (fitness)” (10)

• “Dancing” for manually entered “dancing (latin and ballroom)” (7), and “dance workout” (1)

• “Workout” for manually entered “Exercising” (1)

• “Active socially” for manually entered “chiro geven” (3), and “scouts” (1) (both Dutch names of youth organizations)

The resulting 21 PA types with their corresponding total amount of instances were: Walking (314), Cycling (142), Dancing (35), Running (43), Swimming (2), Cleaning (58), Boxing (4), Power training (8), Football (1), Yoga (20), Weight training (15), Fitness (22), Workout (10), Working (4), Gardening (10), Padel (6), Horseback riding (2), Active socially (4), Stretching (20), Squash (1), and Bootcamp (3).

Appendix C Random forest classifier hyperparameters

The resulting most optimal hyperparameters for all four Random Forest Classifiers were:

• n estimators: 617

• min samples split: 5

• min samples leaf: 10

• max features: ‘auto’

• max depth: 3

• bootstrap: False