Can a Recommender System Support Treatment Personalisation in Digital Mental Health Therapy? A Quantitative Feasibility Assessment Using Data from a Behavioural Activation Therapy App

Abstract

Recommender systems have the potential to improve the user experience of digital mental health apps. Personalised recommendations can help users to identify therapy tasks that they fnd most enjoyable or helpful, thus boosting their engagement with the service and optimising the extent to which it helps them to feel better. Using a dataset containing 23,476 ratings collected from 973 players of a mental health therapy game, this work demonstrates how collaborative fltering algorithms can predict how much a user will beneft from a new therapy task with greater accuracy than a simpler baseline algorithm that predicts the average rating for a task, adjusted for the biases of the specifc user and specifc task. Collaborative fltering algorithms (matrix factorisation and k-nearest neighbour) outperform this baseline with a 6.5-8.3% improvement in mean absolute error (MAE) and context-aware collaborative fltering algorithms (factorisation machines) outperform with a 7.8-8.8% improvement in MAE. These results suggest that recommender systems could be a useful tool for tailoring recommendations of new therapy tasks to a user based on a combination of their past preferences, the ratings of similar users, and their current context. This scalable approach to personalisation -which does not require a human therapist to always be in-the-loop -could play an important role in improving engagement and outcomes in digital mental health therapies.

Understanding What Drives Long-term Engagement in Digital Mental Health Interventions: Secondary Causal Analysis of the Relationship Between Social Networking and Therapy Engagement

“…A recent development in the field is the personalization of digital mental health interventions to enhance the user experience and engagement [3][4][5][6]. These include the use of machine learning algorithms [3,4] and prompts or reminders [5,6] to tailor interventions to individual users needs and patterns of engagement.…”

Health Recommender Systems Development, Usage, and Evaluation from 2010 to 2022: A Scoping Review

“… Government database: four studies used the United States Department of Agriculture (USDA) food composition guidelines [ 25 , 51 , 52 , 76 ]; other studies extracted dietary information from the Taiwan food nutrition database of the Food and Drug Administration [ 38 ], Taiwanese Snacks Nutrition Analysis [ 37 ], Food Composition Database (FCD)”, and Japan Preventive Association of Life-style related Disease (JPALD) [ 78 ]. Other resources: a range of other information resources were utilized for recommended items, including online health community [ 48 , 49 , 50 ], online registered doctors [ 66 , 68 , 70 , 71 ], restaurants [ 40 ], and psychotherapy approaches [ 47 , 74 ] downloaded from Google Play Store. …”

“…Other resources: a range of other information resources were utilized for recommended items, including online health community [ 48 , 49 , 50 ], online registered doctors [ 66 , 68 , 70 , 71 ], restaurants [ 40 ], and psychotherapy approaches [ 47 , 74 ] downloaded from Google Play Store.…”

“…For example, an HRS was compared with health conditions in general or no recommendation scenarios. (2) System evaluation not involving users: 19 studies either directly invited experts to evaluate the system or compared HRS generated recommendations with expert recommendations for the same scenario [ 23 , 24 , 30 , 34 , 35 , 37 , 38 , 39 , 47 , 51 , 52 , 56 , 58 , 59 , 62 , 71 , 72 , 84 , 85 ].…”

“…A variety of metrics were applied to the performance assessment of personalization. (1) forty-six percent of the studies (29/63) used performance criteria (e.g., recall, mean reciprocal rank, normalized discounted cumulative gain, precision, mean absolute error, receiver operating characteristics, and F1) [ 23 , 29 , 34 , 35 , 36 , 37 , 38 , 39 , 41 , 42 , 45 , 47 , 48 , 49 , 50 , 51 , 52 , 58 , 59 , 62 , 63 , 64 , 66 , 68 , 72 , 74 , 75 , 80 , 81 ]. (2) twenty studies used user-centered usability criteria (e.g., usability, satisfaction, acceptance, perceived usefulness) [ 27 , 28 , 30 , 31 , 32 , 33 , 40 , 44 , 55 , 56 , 57 , 60 , 63 , 64 , 65 , 67 , 70 , 79 , 84 , 85 ].…”