We aim to develop a virtual coaching (VC) system which facilitates cancer patients with the formation of positive health habits. The VC should adjust content and delivery of recommendations based on user's preferences. We conducted a survey that elicited potential VC users' perceptions about the appropriate context for notification, aspects driving their motivation, and their ability to perform a selected activity in diferent circumstances. We analyzed 173 survey responses, identified diferent types of users and formulated suggestions how to adjust application content and notification timing to increase their susceptibility to non-medication health interventions.
In this preliminary work, inspired by Ho et al. [
The VC ideally would adjust recommendations based on users’ preferences related to motivation, ability and prompt and to users’ personality traits. We also consider which information about the user is the most informative for adjustment of the VC operations. 3. To map survey responses to personality traits, making a first step toward understanding if personality plays a role in factors impacting habit formation.
We designed a questionnaire consisting of 35 questions: 3 demographic questions (age, gender, nationality), 1 question related to cancer diagnosis, 1 question with a selection from a list of mental well-being exercises that the user would like to perform habitually, 12 closed questions targeting understanding of respondents motivation for performing of health-related activities, 5 closed questions considering factors related to the ability to perform the activity, and 10 closed questions considering the impact of the users’ current context on their responsiveness to prompts and their preferred number of prompts. The questionnaire was concluded with 3 open questions relating to motivation, ability and prompt. In the closed questions, respondents were asked to rate using a 5-point Likert scale their agreement to given statements.
We obtained ethical approval from the University of Haifa’s Ethics Committee and distributed the survey in two languages: English — through the project’s social media and email lists; and Italian — through an Italian Association of Cancer patients (AIMAC). The response to the survey was voluntary and no financial compensation was provided.
173 participants from 19 nationalities responded to our online survey, 98 of whom also replied
to the additional BFI-10 [
For this analysis, we used only the questions from our survey related to Fogg’s three elements of motivation, ability, and prompt. We used the Agglomerated Clustering (AC) and Random Forest (RF) algorithms implemented in scikit-learn for analysis of the survey responses.
The AC was used to hierarchically cluster the respondents based on their rating of each question and the RF was used to identify the questions that help to discriminate between diferent types of respondents. We adopted a four cluster assignment discovered during the initial run of AC as labels for the training of RF classifier, where inputs were responses to each question. We used 75% of responses for training and the remaining 25% for testing.
The RF model trained with balanced class weights and a minimum of 15 samples at a leaf node achieved 72% classification accuracy on the test set. To identify questions that were the most informative in learning to assign the clusters we used two feature importance measures: Mean Decrease in Impurity (MDI) and Mean Decrease in Accuracy (MDA). Using each method, we selected the top 9 questions and then took their intersections, which resulted in 5 questions that were identified by both methods as important.
Finally, we ran AC again with a number of clusters fixed to 4, in order to cluster the responses
only based on the top 5 questions. The resulting number of responses in each cluster was: 21,
41, 55, and 56 (Fig. 2a). The clustering was run 10 times with 10% of respondents left out to
verify the stability of these new clusters. In each run, we used the adjusted Rand index [
To understand how respondent types (clusters) difer and why sometimes cluster assignment
may change, we visualized the distribution of responses per cluster for the top 5 questions (Fig.
2). For better interpretability, the responses are grouped into three categories: 1–2 (”no”), 3
(”neutral”), 4–5 (”yes”). It seems that the diference between clusters 0 and 1 vs 2 and 3 is in terms
of motivation drivers. Respondents in the former two clusters would not be motivated by the
introduction of the social component of competition (Fig. 2b and 2d). Most of the respondents
from cluster 0 would also not be driven by their self-improvement (Fig. 2c), whereas including a
progress-tracking in notification content or app display might motivate respondents in clusters
2 and 3. The diference between clusters 2 and 3 arises when considering situations in which
the users might respond to prompts (Fig. 2e and 2f). The users in cluster 3 might be more
responsive than those in cluster 2. Interestingly respondents in clusters 1 and 2 were confident
that they would not respond to activity recommendations when they are stressed (Fig. 2e). This
suggests that just-in-time stress management interventions, such as the recommendation of
deep-breathing exercises, might be inefective for those groups of users. It might be important
to know this prior to triggering the intervention, such that these users should not be prompted
at times of stress (e.g., as detected from the blood volume pulse measured by a smartwatch [
We investigated whether the discovered clusters difer in terms of the respondents’ personality traits. For this analysis, we kept only the respondents who also completed the BFI-10 personality questionnaire. A one-way ANOVA revealed that there was no significant diference in Extraversion, Agreeableness, Conscientiousness, Neuroticism, or Openness between the four clusters ( = [1.13, 1.57, 0.48, 0.61, 0.97], - = [0.34, 0.19, 0.69, 0.60, 0.41]). We also trained Extra Tree Regressor to predict each personality trait given the survey question ratings. The model was trained on 75% of responses and tested on 25%. The mean absolute prediction error was 1.6, 1.2, 1.3, 1.7, and 1.5 for Extraversion, Agreeableness, Conscientiousness, Neuroticism and Openness, respectively. Responses to questions were generally not linked to personality traits. One question related to mood and motivation for spontaneous action (Fig. 3), was related to neuroticism, suggesting that for a person who scores highly on neurotic trait, recommendation of new activity might be inefective when they are experiencing lower emotional valence.
We conducted a survey to understand the diference between factors afecting people’s motivation to perform health-related activities and the context in which they are able to respond to VC recommendations. Encouragingly, the majority of the respondents were open to receiving at least one prompt daily and 111 of the 173 respondents’ motivation might be spiked by including gamification components such as score dashboard for comparison with peers or personal score tracking.
Interestingly, the person’s internal state of ’stressed’, ’tired’, or ’low mood’ resulted in higher perceived responsiveness to prompts than the person’s location or motion. We also investigated if personality traits are related to identified respondents clusters and factors impacting the formation of new habitual behavior. Other than the association between neurotic trait and motivation for novel action in a low mood context, we found no links.
The gathered survey responses and presented user clustering might drive the simulation of people responsiveness to intervention in various contexts. However, the survey results are insuficient to conclude that personality and habit formation are unrelated. The hypothetical self-reported responses to behavior intervention might vary from the actual response and hence the potential link or lack of it must be evaluated experimentally. The limitations of this work also include relatively small number of responses compared to complexity of the habit formation problem and potential selection bias due to distribution of the survey through social and professional network of the authors.
The CAPABLE project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 875052.