Towards Explainable AI Modelling on Brain Ageing

Zhang, Zhaonian and Jiang, Richard and Williams, Bryan (2025) Towards Explainable AI Modelling on Brain Ageing. PhD thesis, Lancaster University.

Text (2025ZhaonianZhangPhD)
2025ZhaonianZhangPhD.pdf - Published Version
Download (60MB)

Abstract

Machine learning, when combined with advanced neuroimaging such as three-dimensional Magnetic Resonance Imaging (MRI), has opened new possibilities for understanding brain health and disease. Among MRI modalities—structural MRI (sMRI), functional MRI (fMRI), and Diffusion Tensor Imaging (DTI)—sMRI is most widely applied in machine learning research, as it provides detailed measures of cortical thickness, gray matter volume, and subcortical anatomy. Despite significant progress, existing brain age estimation methods face three persistent challenges: limited predictive accuracy across diverse age groups, insufficient interpretability of model predictions, and a lack of fairness in mitigating demographic biases such as agerelated bias. These gaps restrict the utility of brain age as a reliable biomarker in both research and clinical settings. This thesis addresses these limitations by developing new approaches for brain age estimation from sMRI, aiming to improve accuracy, enhance interpretability, and incorporate fairness. To this end, I compiled several large-scale sMRI datasets and proposed three models: the Nonlinear Age-Adaptive Ensemble (nl-AAE), the Triamese Vision Transformer (Triamese-ViT), and the Democratic AI framework (u-DemAI). The nl-AAE improves predictive accuracy by dynamically weighting multiple base learners according to age groups, achieving a mean absolute error (MAE) of 3.19 years (r = 0.95). The Triamese-ViT leverages three orthogonal MRI views and integrates built-in interpretability, meaning that its attention mechanisms generate explanatory maps directly during the prediction process. These intrinsic explanations, validated against conventional explainable AI (XAI) techniques, highlight age-related and ASD-related brain regions consistent with established clinical findings. The u-DemAI framework extends beyond predictive performance by incorporating user personalization into the framework, enabling community-driven model updates and explicitly addressing fairness—particularly reducing age-related bias (ageism) in predictions. Taken together, these contributions advance the state of the art in brain age estimation by combining accuracy, interpretability, and fairness. More broadly, this work demonstrates how democratic principles can be embedded into machine learning frameworks to promote equitable, transparent, and socially responsible applications in neuroscience and clinical practice.