Towards an Empathetic and Meta-Cognitive Digital Therapy Framework

Abstract

We present an early design of Mindful-AI, a digital therapy framework that perceives multimodal affect, plans Cognitive Behavioural Therapy (CBT) strategy and self-monitors its own language. The proposed model architecture consists of a state vector incorporating sentiment, Electroencephalogram (EEG) band-power, heart-rate variability (HRV) and galvanic skin response (GSR), a reinforcement learning (RL) planner that learns to probe, reflect or reframe, a large language model (LLM) that generates the chosen act in natural language and a critic either accepts, regenerates or escalates the reply. Preliminary results suggest the potential of the proposed therapy agent in real world settings.