REPOSITORY > RESULTS

Abstract

The increasing prevalence of mental health issues worldwide has amplified the significance
of computational psychotherapy, which includes creating computational tools for
the mental healthcare and tools to support existing mental health professionals. This
work presents a computational psychotherapy system that predicts and forecasts mental
health issues in users, and utilizes a conversational agent to induce behavior and attitude
change. The thesis centers around two main contributions. The first contribution is a
novel, golden standard dataset, which includes panel data, encompassing multiple individuals
at multiple time intervals. It incorporates 1495 instances of quantified stress, anxiety,
and depression levels, as well as symptom scores derived from diagnostic-level questionnaires,
accompanied by qualitative daily diary entries. The second contribution is the
system itself. The hypothesis posits that for the system to be effective in inducing mental
health issues relief with a conversational agent, it needs to simulate theory of mind - the
cognitive ability to understand others and act accordingly. The system simulates theory
of mind with an artificial cognitive architecture comprised of an ensemble of computational
models. It uses psychological as well as cognitive modelling and machine learning
models trained on the novel dataset, all in conjuction with novel domain ontologies. The
system was evaluated through a computational experiment on mental health phenomena
prediction and forecast from quantitative scores and qualitative text diary entries, and an
empirical interventional study on relieving mental health issues in participants where it
was compared against Woebot. The latter system was chosen as it is currently the most
cited freely available system with the most replicated positive outcomes. This work’s system
showcased superior performance compared to state-of-the-art systems in terms of both
the number of detected mental health categories and detection accuracy. It achieved an
accuracy of 91.41% using the kNN algorithm (chosen for its explainability, despite several
other algorithms performing slightly better), surpassing the highest accuracy of one of the
other systems which reached 84% using Long short-term memory. The highest accuracy
for 7-day forecasting achieved 87.68%, while other systems were not able to forecast trends.
In the empirical interventional study on 42 participants in a simulated daily check-in, the
system outperformed Woebot in reducing stress (p = 0.048) and anxiety (p = 0.040) levels
in participants, while both failed to reduce their depression levels (p = 0.688). With
confirmed hypothesis, it was evaluated that this system performs on par or better than
comparable state-of-the-art systems.