Designing an Intelligent Cognitive Assistant as Persuasive Technology for Stress, Anxiety and Depression Relief

Tine Kolenik, Martin Gjoreski, Matjaz Gams

Introduction

Stress, anxiety, and depression (SAD) represent a significant and growing burden on global healthcare systems. While traditional psychotherapy is highly effective, its reach is often limited by high costs, a lack of 24/7 availability, and the persistent social stigma associated with seeking help. Intelligent Cognitive Assistants (ICAs)—chatbots designed for psychotherapeutic support—have emerged as a scalable, always-on alternative. However, current state-of-the-art systems often fail to fully leverage the potential of Persuasive Technology. This paper proposes a novel Psychotherapeutic Intelligent Cognitive Assistant (PICA) architecture designed to bridge the gap between automated support and personalized, data-driven intervention.

The Limitations of Current ICAs

Existing mental health chatbots, such as Woebot or Tess, typically rely on simple decision trees or basic emotion detection from text. These systems often lack a deep "Theory of Mind" (ToM)—the ability to model and understand a user's unique psychological state. Furthermore, they rarely integrate established behavior change theories or adapt to real-time physiological data, relying instead on manual user reporting which can be inconsistent or subject to bias.

Proposed Methodology: The PICA Architecture

The authors propose a PICA framework centered on a domain-specific Theory of Mind, utilizing several key methodological components to enhance therapeutic efficacy:

  • Deep User Modeling: The system constructs a comprehensive profile based on the Big Five personality traits (gathered via dialogic questionnaires) and clinical SAD scores tracked through the DASS-21 questionnaire.
  • Multimodal State Tracking: To bypass the friction of constant self-reporting, the PICA integrates with a smart bracelet. This allows for continuous, real-time monitoring of biophysiological measurements, providing an objective view of the user's stress and anxiety levels.
  • Simplified Interaction Design: The interface utilizes a primarily button-based interaction model rather than open-ended natural language. This reduces NLP complexity and ensures the system's behavior remains predictable and clinically safe.
  • Expert Ontologies: The bot’s domain knowledge is structured using custom ontologies rooted in cognitive and behavioral sciences, specifically focusing on Behavior Change (BC) and SAD relief.

Reinforcement Learning and Personalized Persuasion

A core innovation of the PICA is the application of Reinforcement Learning (RL) to strategy selection. Rather than following a static script, the system treats therapy as an optimization problem. By evaluating historical interactions, the RL agent learns which specific persuasive strategies and "nudges" are most effective for a specific user’s personality and current mental state. This allows the assistant to tailor the style and framing of its messaging, theoretically increasing the likelihood of successful behavior change.

Key Insights

  • Personality-Driven Intervention: By modeling Big Five traits, the assistant can move beyond one-size-fits-all advice, framing interventions in a way that resonates with the individual user's psychological makeup.
  • Proactive Nudging: The integration of wearable data allows the system to intervene at the exact moment physiological stress is detected, rather than waiting for the user to initiate a session.
  • Dynamic Learning: Framing symptom reduction as a reward signal for an RL agent allows the chatbot to dynamically improve its therapeutic approach over time.

Future Directions

As this is currently a design proposal, the primary limitation is the lack of empirical validation in a clinical setting. The authors have established a quantitative benchmark based on previous state-of-the-art systems, which achieved roughly 15% to 20% improvements in user mood.

Future work will focus on the full implementation of the PICA architecture and the execution of a randomized controlled trial. The authors intend to replicate the 2017 Woebot study by Fitzgerald et al., replacing the legacy system with the PICA to directly measure the efficacy of their Theory of Mind and wearable-integrated approach against current industry standards.

Summary updated

BibTeX
@inproceedings{kolenik2020designing,
  author = {Tine Kolenik and Martin Gjoreski and Matjaz Gams},
  title = {{Designing an Intelligent Cognitive Assistant as Persuasive Technology for Stress, Anxiety and Depression Relief}},
  booktitle = {PERSUASIVE (Adjunct)},
  year = {2020},
}
Figure