D4 – Development of Smart Coach

 
 

Project team

Project PI – David Reinkensmeyer

Sangjoon Kim, George Collier, Amanda Rabinowitz, Veronica Swanson, Danny Zondervan

Purpose/aims

Studies have shown the positive effects of conversational agents in multiple healthcare domains such as physical activity, weight loss, alcoholism treatment and management of mental health conditions. However, chatbots have not yet been developed and tested to encourage rehabilitation exercises at home in individuals recovering from a stroke.

 The goal of this project is to develop and validate a conversational agent (called “Smart Coach”) to increase user home exercise adherence following stroke. Smart Coach will help individuals establish and maintain the long-term exercise habits that are crucial for their recovery. Smart Coach will use a conversational, SMS-based (e.g., text messages, Facebook Messenger) interface to help users establish a user-tailored workout plan, provide exercise reminders and encouragement, and track progress, leveraging the real-time exercise data sensed by FitMi.

 The coaching rules (i.e., intelligence) of Smart Coach will be based on the literature of habit formation/behavioral psychology and previous studies of FitMi. The coaching rule will be iterated through a user-centered design process, performing focus groups and feedback interviews to further optimize message content, timing, and frequency.

Status

The groundwork for initial coaching rules (i.e., coaching intelligence) has been established. A prototype of Smart Coach is under development. Focus groups and feedback interviews to optimize message content, timing, and frequency have been conceptualized.

Key accomplishments

  • We analyzed 2,581 commercial users of FitMi who used the system for at least 20 weeks to investigate the impact of the initial exercise frequency on longer-term engagement with FitMi. We observed that users who exercised at least 4 times per week during the initial 6 weeks of FitMi usage had higher and more stable habit strengths over 20 weeks. These results are consistent with previous work on exercise habit formation for unimpaired populations (Gardner, 2015). We presented these results at the DARE 2023 conference. Kim SJ, Swanson VA, Collier GH, Rabinowitz AR, Zondervan DK, Reinkensmeyer DJ (2023) Modeling the relationship between initial exercise frequency and longer-term engagement with a sensorized home rehabilitation exercise program.

  • We have established initial coaching rules that the prototype Smart Coach will follow when generating conversational prompts. These coaching rules are based on our previous results showing that 1) initial exercise frequency, 2) user impairment levels, 3) initial success, and 4) steadiness of use – are associated with persevering at-home rehabilitation exercise.

  • We have developed and prototyped a powerful and flexible architecture based on our experience with multiple years of development of chatbot systems (Rabinowitz et al., 2022).

 The architecture of this system is designed to achieve the following goals:

  • To be flexible and extensible.

  • Support loose coupling between significant system components to reduce dependencies, allow substitution of principal components as desired, and leverage available cloud services.

  • Provide scalability, availability, and security.

  • Leverage cloud computing to reduce costs and support growth as needed.

  • To support multiple approaches to providing the "intelligence" of the coaching system.

Figure 1: Smart Coach system architecture diagram

  • Smart Coach Control System (item 2 in Figure 1): The heart of the Smart Coach is a custom data-driven application that implements the "coaching" elements of the system. This system currently is a running alpha prototype implemented in Python. As the "brains" of the Smart Coach, the controller has the following key functionalities:

    • Harvesting, modeling, and storing participant data from the FitMi application. The FitMi participant data will be held in a robust, secure store. In future system versions, participant data from other rehabilitation systems will be collected and leveraged.

    • Implementing the "smart" coaching component of the system. The Smart Coach intelligence module will be a "pluggable" component allowing for different approaches to generating advice and guidance for a participant. The team has discussed, and prototyped approaches drawn from fields as diverse as machine learning and behavioral change theory. However, our initial approach will be a rule-driven engine that leverages habit formation theory and statistical analysis of historical FitMi game use (Ramos Muñoz et al., 2022).

    • Interfacing into the chat dialog portion of the system. We are implementing the chat dialog management system in Power Virtual Agents (Intelligent Virtual Agents and Bots | Microsoft Power Virtual Agents, n.d.). Power Virtual Agents is a Microsoft chatbot development and deployment framework that runs in the Azure cloud. Power Virtual Agents have machine-learning-based natural language understanding capabilities. Power Virtual Agents can take advantage of the capabilities of ChatGPT (PVA ChatGPT Release, 2023).

  •  We recently submitted an SBIR Phase II grant to support the feasibility test of Smart Coach as a subscription service add-on to FitMi.