Evaluating Classical and Novel Machine Learning Approaches for Early Alzheimer's Detection Using fMRI Data

Identifier:  TFM:2521

Handle: https://hdl.handle.net/20.500.11797/TFM2521

Authors:  Baidal Marco, Miguel

Abstract:
Alzheimer’s disease (AD) is characterized by early biological alterations that begin many years before the appearance of clinical symptoms, making early detection both challenging and crucial for effective prediction and intervention. In this study, a systematic evaluation of some predictive models is applied with the aim to assess the effectiveness of functional MRI connectivity and amplitude features for identifying participants at risk of AD. This analysis compares both traditional machine learning techniques and modern neural network architectures, evaluating their performance in terms of accuracy, but also in other dimensions such as dimensionality reduction, feature selection, and the handling of class imbalance. Moreover, explainability methods to better understand which features are more important in prediction are applied, providing relevant insights into the mechanisms of the disease. The transferability of the findings is further tested using an independent validation dataset. The results suggest that, despite the promise of deep learning, classical approaches remain highly competitive, highlighting the complexity of early AD prediction.