Repository logo
 

Multimodal, Personalized and Explainable AI in Mental Health: Early Diagnosis and Prognosis using Longitudinal Data

Files

Thesis(12.9 MB)

Date

2025

Authors

Budhraja, Sugam

Supervisor

Doborjeh, Maryam
Goh, Wilson
Kasabov, Nikola

Item type

Thesis

Degree name

Doctor of Philosophy

Journal Title

Journal ISSN

Volume Title

Publisher

Auckland University of Technology

Abstract

The early diagnosis and prognosis of mental health disorders remain formidable challenges, often complicated by their multifaceted origins and evolving presentation. This thesis tackles these challenges by developing explainable artificial intelligence (AI) models that combine multimodal and longitudinal data—spanning clinical, cognitive, genetic, and social domains—to improve the accuracy, transparency, and real-world usability of mental health predictions. The work introduces several novel methods, including the Filter and Wrapper Stacking Ensemble (FWSE) for robust biomarker discovery in high-dimensional gene expression data, and the Dynamic Attention Gateway (DAG) for Liquid State Machines, which brings both interpretability and temporal sensitivity to time-series analysis. In integrating diverse data sources, the Mosaic Liquid State Machine (Mosaic LSM) architecture demonstrates that fusing clinical, neurocognitive, genetic, and social data can yield predictive models that outperform single-modality approaches in identifying individuals at ultra-high risk for mental illness. Empirical results across real-world datasets—including the longitudinal LYRIKS study—show that these models not only match or exceed the accuracy of standard machine learning approaches but also offer clinicians transparent, case-by-case explanations for their predictions. For example, using pathway-level gene aggregation and attention-based mechanisms, prediction accuracies above 95% were achieved for certain high-risk mental health classifications, while simultaneously surfacing biologically meaningful markers that align with established literature. To promote adoption beyond the data science community, this research also delivers NeuroGeMS, an open-source GUI software that makes advanced multimodal AI accessible for researchers in the biomedical domain. Taken together, these contributions lay groundwork for personalized and explainable AI in mental health.

Description

Keywords

Source

DOI

Publisher's version

Rights statement

Permanent link

http://hdl.handle.net/10292/19243

Collections

Doctoral Theses