Skip to main content
ABCT 2023 Annual Convention Main banner
Final Program


Program Flipbook
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.
Presentation Icons
Ticketed Session
Ticketed Session
Recorded Session
Recorded Session
Back
ViewAttendees 5

Symposia

Technology

Session: Using machine learning in scalable, adaptive digital mental health interventions

3 - (SYM 99) Using Contextual Bandits to Tailor Text Messages in an Adaptive Intervention for Depression and Anxiety Symptom Self-management

Saturday, November 18, 2023
4:15 PM – 5:45 PM PST
Location: Chiliwack (402), Level 4

Although digital mental health interventions offer the hope of increasing access to evidence-based interventions through scalable, low-cost programs, a major challenge is how to tailor an automated intervention to a given person’s needs and preferences over the course of treatment, a process that happens naturally in face-to-face treatment. Adaptive interventions aim to address this problem with methods including branching logic, natural language processing, and reinforcement learning algorithms. Here, we present one example of tailoring a digital mental health intervention with contextual bandits, a simpler form of reinforcement learning. Small Steps is an automated 8-week text message intervention that supports management of depression and anxiety symptoms through self-experimentation with a diverse set of cognitive, emotional, and behavioral strategies. On each day of the program, users may receive messages with psychoeducational context, first-person stories from peers, encouragement to write their own messages to peers, or “modular dialogues,” a set of brief interactions determined either by randomization or several contextual bandit algorithms. These algorithms aim to maximize variables related to engagement and liking by learning how users responded to similar dialogues in similar contexts, with contextual variables including things like their recent activity with the system, time of day, day of week, day of the study, user’s mood and energy, user’s symptom severity, and user’s average responsiveness to the system. These contextual bandits are used to determine variables about the framing, content, type, and timing of messages. We will describe the process of iteratively refining the contextual bandit algorithms, including decision-making around selection and treatment of contextual and reward variables, along with preliminary results comparing user responses to the bandit-determined messages versus the randomly selected messages.