TY - DATA T1 - Collaboratively Setting Daily Step Goals with a Virtual Coach: Using Reinforcement Learning to Personalize Initial Proposals - Data and Analysis Code PY - 2024/01/23 AU - Martin Dierikx AU - Nele Albers AU - Bouke L. Scheltinga AU - Willem-Paul Brinkman UR - DO - 10.4121/53f2d238-77fc-4045-89a9-fb7fa2871f1d.v1 KW - behavior change KW - physical activity KW - goal-setting KW - conversational agent KW - chatbot KW - virtual coach KW - reinforcement learning N2 -

This is the data and analysis code underlying the paper "Collaboratively Setting Daily Step Goals with a Virtual Coach: Using Reinforcement Learning to Personalize Initial Proposals" by Martin Dierikx, Nele Albers, Bouke L. Scheltinga, and Willem-Paul Brinkman. The paper develops a dialog to collaboratively set daily step goals with a virtual coach and analyzes the use of reinforcement learning to personalize the initial step goal proposal in the dialog.


Study

The paper is based on data collected from a study conducted in June and July 2023 for the publicly available Master's thesis by Martin Dierikx (http://resolver.tudelft.nl/uuid:4f2c12de-9b9f-4e3f-ad3a-902947d693bb). In this study, 235 people were invited to between one and five conversational sessions with the text-based virtual coach Steph. In each session, Steph asked questions to determine people's current state based on their mood, sleep quality, available time, motivation, and self-efficacy. Afterward, Steph calculated a recommended daily step goal based on the user's previous walking behavior. Based on this recommended goal, Steph gave users three initial goal options, each 100 steps apart. Thereby, the options were randomly changed in one of five possible ways: 1) decrease by 400 steps, 2) decrease by 200 steps, 3) keep the same, 4) increase by 200 steps, or 5) increase by 400 steps. Users could select one of the presented goal options as well as indicate that they wanted a different goal. The next session started by asking users about the number of steps they took on the previous day. Data collected from this study was used to fit and analyze a reinforcement learning model for choosing initial step goal proposals.


The study was pre-registered in the Open Science Framework (OSF): https://doi.org/10.17605/OSF.IO/6JQPK.


The Human Research Ethics Committee of Delft University of Technology approved our study (Letter of Approval number: 3016).


Links to further resources:

  1. The Rasa-based implementation of the virtual coach Steph is available here: https://doi.org/10.5281/zenodo.8382413.
  2. A video of a dialog with the virtual coach is available here: https://youtu.be/FSpG-G0zc-o.


Data

We collected data in several study components:

  1. Demographic data collected from participants' Prolific profiles (e.g., age, gender).
  2. Data collected from a prescreening questionnaire (e.g., Godin leisure-time physical activity).
  3. Data collected during the conversational sessions (e.g., mood, number of steps taken on the previous day).
  4. Data from the post-questionnaire (e.g., how personal the goals felt to participants, how difficult it was to reach the goals).



If you have any questions, please contact Nele Albers (n.albers@tudelft.nl) or Willem-Paul Brinkman (w.p.brinkman@tudelft.nl).

ER -