Modeling Health Coaching Dialogues for Behavioral Goal Extraction

Abstract

Health coaching has been identified as a successful method for motivating and maintaining health behavior changes. Unfortunately, personal health coaching is time-and resourceintensive, and cannot scale up. Previous research focuses on developing conversational systems that can provide automated coaching to patients. But most of these systems rely on a predefined set of input/output mappings, focus more on general goal setting, and do not provide follow-up during goal implementation. Therefore, my research contributes to building a virtual assistant health coach to help patients set and achieve a Specific, Measurable, Attainable, Realistic, and Time-bound (SMART) goal.The first step towards building a successful domain-specific dialogue system is to understand the users' needs and the environment in which it will be used. Since no other public health coaching dialogue data exists, we collected data that I annotated for a variety of features of interest. These annotations include domain-specific slots called SMART attributes at the word level, the intent behind a participant's text message called dialogue acts, and higher-level portions of the dialogue performing a specific function called stages and phases. I then used these annotations to build the language understanding component of the dialogue system and design the goal extraction pipeline, which extracts the physical activity goal the user and coach negotiate via text messages. I proposed a two-step goal extraction process that uses phases and dialogue acts to identify the correct SMART goal attributes and show that dialogue acts help more than the higher-level conversation flow captured by phases. These goal summaries will xv SUMMARY (Continued) help the health coaches to easily recall patients' past goals and use them to suggest a realistic future goal. Finally, I assessed the usability of the generated goal summaries with the help of three health coaches in an offline setting.

Artificial Intelligence in Rehabilitation Targeting the Participation of Children and Youth With Disabilities: Scoping Review

“…The remaining remotely delivered interventions commonly apply ML and NLP, potentially indicating the suitability of ML and NLP for use in remote pediatric rehabilitation interventions using AI. ML and NLP have been used in a range of health interventions to promote behavioral changes, such as physical activity and healthy diet, including goal-setting [ 136 , 137 ]. Given the existing evidence on the use of AI for goal-setting in other health care domains [ 136 , 137 ] and the importance of gaining efficiency in enacting the complex process of goal-setting in pediatric rehabilitation [ 138 , 139 ], the lack of attention to goal-setting in this review indicates a clear knowledge gap warranting future research.…”

“…ML and NLP have been used in a range of health interventions to promote behavioral changes, such as physical activity and healthy diet, including goal-setting [ 136 , 137 ]. Given the existing evidence on the use of AI for goal-setting in other health care domains [ 136 , 137 ] and the importance of gaining efficiency in enacting the complex process of goal-setting in pediatric rehabilitation [ 138 , 139 ], the lack of attention to goal-setting in this review indicates a clear knowledge gap warranting future research. Emerging electronic participation–focused interventions such as the Participation and Environment Measure–Plus [ 140 - 143 ] with individual goal-setting as an integral part of their intervention might benefit from exploring the use of AI to fill this knowledge gap.…”

Using Fitbit as an mHealth Intervention Tool to Promote Physical Activity: Potential Challenges and Solutions

“…Regardless of the type of issue faced during the intervention period, we learned a single strategy that helped to address most of the challenges during the intervention delivery: devoting time to in-person intervention meetings or placing telephone calls when necessary to discuss potential challenges or issues associated with the use of Fitbit. Across the studies we conducted, we dedicated a portion of the in-person meeting to Fitbit-related content [ 9 , 10 , 20 , 35 ]. During this time, participants were free to externalize any concerns or issues related to Fitbit use, and research staff were ready to assist with possible problems or questions.…”

Summarizing Behavioral Change Goals from SMS Exchanges to Support Health Coaches

“…We decided to use the CRF model with an F1 macro score of 0.81. In our previous work (Gupta et al, 2020b), we used models with word2vec embeddings (Mikolov et al, 2013) but found ELMo embeddings to perform better.…”

“…We previously showed in Gupta et al (2020b) that metrics like BLEU (Papineni et al, 2002) and ROUGE (Lin, 2004) are not appropriate for our extraction-based goal summaries as they are sensitive to exact word match (Reiter, 2018). That is, if a given word, say 'two', is classified as days number instead of distance, they will still output a high score as 'two' is in the reference summary.…”