Skip to main content

William Kearns

Graduated: March 17, 2023

Thesis/Dissertation Title:

Enhancing Empathy in Text-Based Teletherapy through Emotional State Inference

Over half of the U.S. population lives in an area without adequate access to mental health care and the unmet demand for mental health services has shifted to care providers who have not been trained to provide mental health support. This work represents a step toward addressing this supply-demand imbalance by applying recent advances in conversational AI.
The central hypothesis of this work is that both the quality and efficiency of text-based teletherapy can be improved through conversational AI. This was evaluated using mixed-methods approaches with three aims: (1) I explored the ability of computational methods to infer high-fidelity representations of self-reported emotional states, (2) I evaluated these representations as features to predict empathetic responses, (3) I piloted this system as a component of an AI-augmented teletherapy platform for the delivery of problem-solving therapy by nurses and psychologists.
(1) Prior to this work, emotion recognition from conversation (ERC) methods had only been tested on crowdsourced data labels that (a) were inferred by annotators rather than self-described, and (b) did not cover the breadth of emotional states experienced as a result of daily events. This prior work was insufficient to assess the applicability of these methods to characterize self-reported emotional states in the context of check-ins where the emotional states may not be explicitly expressed. To address this gap, I evaluated emotion detection and emotional state inference methods on event-emotional state pairs collected through a daily journaling exercise delivered by SMS. I found that emotional state inference methods improved performance on the task of predicting reported emotions by 71.3% relative to emotion detection methods.
(2) The messages from the daily journaling exercise were labeled by experts based on how they would respond empathetically to them in the context of teletherapy. I found that the addition of emotional state inferences to these messages improved the performance of models on the task of predicting these labels, which in turn indicate appropriate expert-authored empathetic responses to a given utterance.
(3) Quantitative results of the AI-augmented provider platform indicate that the system decreased response times by (+29.34%**; p=0.002), tripled empathetic response accuracy (+200%***; p=0.0001), and increased goal recommendation accuracy (+66.67%**; p=0.001). Structured qualitative interviews indicated that the care providers who used the system felt it would make providing therapy more efficient, lower cognitive load, and be accessible to care providers without mental health training.

Committee:

Drs. Trevor Cohen, Gina-Anne Levow, Alex Marin, Weichao Yuwen