Sex Differences in Machine Learning Models for Multimodal Biomarker-Based Diagnosis and Prognosis of Alzheimer’s Disease

Abstract

Alzheimer’s disease (AD) impacts the brain in multiple ways and has varied clinical features and progression. Sex differences exist in the epidemiology and fundamental biology of AD. Machine learning (ML) tools can integrate and analyze multiple biomarkers, improving AD diagnosis and prognosis. However, machine learning tools underuse and under-explore predictive modeling of sex differences. This translational narrative review synthesises the existing literature on studies that explore sex differences in Machine Learning models that utilise data collection techniques that incorporate biomarkers from Positron Emission Tomography (PET) scans, Magnetic Resonance Imaging (MRI) scans, Cerebrospinal Fluid (CSF), serum assays, genetic variants, including single-nucleotide polymorphisms (SNPs), and Electroencephalography (EEG). We emphasize the role of sex in biomarker manifestations which include but are not limited to: the dissimilarity of amyloid and tau in PET imaging, the effects of sex in structural and functional brain alterations in MRI, sex-specific alterations in cerebrospinal fluid (CSF) composition (e.g., Aβ42 concentration, total tau, and phospho-tau) and the alterations in peripheral and systemic inflammatory and metabolic blood markers, the modulation of genetic risk particularly in relation to APOE and some of the single nucleotide polymorphisms (SNPs), and the variability in sex differences in EEG network disruption and/or modularity. When sex is incorporated as a biological variable or when sex-stratified models are created, new ML methods, such as deep learning, ensemble techniques, and multimodal fusion frameworks, demonstrate promising diagnostic and prognostic capabilities. Many studies, however, remain constrained by modest datasets, a lack of external validation, and limited interpretability, thereby limiting clinical applicability. This review highlights the importance of sex-aware modeling techniques to advance precision medicine in AD. We offer several methodological suggestions, including analysing disaggregated sex data, aligning data across multiple approaches, using explainable AI, and including longitudinal data. Filling these gaps may help develop ML tools that offer accurate, fair, and clinically useful early detection and prognosis of AD across a wide range of populations.