Neurodegenerative diseases do not happen overnight. They unfold over decades in a complex cascade of molecular, structural, and cognitive shifts. By the time clinical symptoms like memory loss manifest, significant and irreversible brain atrophy has already occurred. To intervene effectively, we need to understand the timeline: when do specific brain regions begin to degenerate, how do different pathologies interact, and why do some individuals progress faster than others?
Historically, tracking this progression was limited to static, single-modality snapshots. However, a major paradigm shift has emerged. By treating disease progression as a continuous, multi-dimensional trajectory, researchers are leveraging large-scale longitudinal datasets to build models that can disentangle normal aging from pathology, map disease subtypes, and predict clinical conversion years in advance.
Disentangling Disease Signal from Normal Aging and Bias
The first challenge in modeling neurodegeneration is isolation. The brain changes naturally as we age; ventricles expand, and cortical thickness decreases. How do we distinguish the benign signature of a healthy 70-year-old brain from the early stages of a neurodegenerative disease?
To solve this, Pennec et al. developed a computational framework to mathematically separate normal aging from disease-specific structural alterations in structural MRI (Disentangling normal aging from Alzheimer's disease in structural magnetic resonance images). By establishing a reference trajectory for healthy aging, the model isolates the specific "deviation vector" caused by pathology.
Beyond biological aging, technical artifacts introduce significant noise. MRI scans collected across different sites often suffer from procedural biases due to variations in scanner field strengths or preprocessing pipelines. Sun et al. addressed this by introducing GSplit LBI, a dual-task learning algorithm designed to explicitly decouple disease-related lesion features from procedural biases (GSplit LBI: Taming the Procedural Bias in Neuroimaging for Disease Prediction). By cleaning the imaging signal of these confounding variables, models can more reliably capture the true biological trajectory of the disease.
Multimodal Fusion and Cross-Modality Translation
No single biological marker tells the whole story. Structural MRI captures late-stage macrostructural atrophy, Positron Emission Tomography (PET) tracks early-stage protein aggregation (such as amyloid-beta and tau), and blood-based biomarkers reveal immediate metabolic or inflammatory changes. To build a holistic progression model, these disparate data streams must be fused.
For instance, Kim et al. demonstrated that combining amyloid PET and structural MRI shape features yields a significantly stronger predictor of clinical decline than either modality alone (Prediction of Conversion from Mild Cognitive Impairment to Alzheimer’s Disease Using Amyloid PET and Brain MR Imaging Data: A 48-Month Follow-Up Analysis of the Alzheimer’s Disease Neuroimaging Initiative Cohort).
However, acquiring multimodal data for every patient is prohibitively expensive and clinically impractical. PET scans are costly and involve radiation, whereas MRIs and blood draws are far more accessible. To circumvent this limitation, researchers use generative AI to perform cross-modality translation. Chen et al. introduced Plasma CycleGAN, a generative model that synthesizes amyloid PET images directly from a patient's T1-weighted MRI and blood-based biomarkers (Plasma CycleGAN: Integrating Blood‐based Biomarkers for Cross‐modality Translation from MRI to PET). By conditioning the image-to-image translation on molecular blood markers, the network infers the underlying biochemical state of the brain, generating highly accurate synthetic PET scans without requiring the actual expensive imaging procedure.
Uncovering Heterogeneity via Subtyping and Staging
A major pitfall of early progression models was the assumption of a single, universal pathway for disease progression. In reality, neurodegeneration is highly heterogeneous. One patient might experience early atrophy in the temporal lobe, while another shows a frontal-dominant pattern.
To map this variation without relying on arbitrary clinical labels, unsupervised machine learning frameworks like Subtype and Stage Inference (SuStaIn) have been developed. The core intuition behind SuStaIn is to simultaneously solve two problems: clustering patients into distinct spatial subtypes and ordering the sequential stages of regional brain volume loss within each subtype.
The authors of recent studies have extended this framework to handle the real-world challenge of missing data. Estarellas et al. applied a missing-data-enabled version of SuStaIn to a large multimodal dataset, successfully identifying five distinct disease subtypes, each characterized by a unique spatiotemporal trajectory of neuroimaging, cerebrospinal fluid, and cognitive markers (Multimodal subtypes identified in Alzheimer’s Disease Neuroimaging Initiative participants by missing-data-enabled subtype and stage inference).
This capability to map fine-grained spatiotemporal trajectories extends across the spectrum of neurodegenerative conditions. Young et al. utilized SuStaIn to uncover how specific genetic mutations dictate distinct patterns of brain atrophy and clinical symptoms over time in frontotemporal dementia (Characterizing the Clinical Features and Atrophy Patterns of MAPT -Related Frontotemporal Dementia With Disease Progression Modeling). Similarly, Scotton et al. leveraged the algorithm to reveal two distinct spatiotemporal subtyping patterns in progressive supranuclear palsy (Uncovering spatiotemporal patterns of atrophy in progressive supranuclear palsy using unsupervised machine learning), demonstrating that what once appeared to be clinical noise is actually highly structured, predictable variation.
Predicting Proximity to Clinical Conversion
From a clinical standpoint, the most critical question for a patient with Mild Cognitive Impairment (MCI) is: When will the disease progress to dementia?
Traditional classification models treat this as a binary problem (e.g., stable MCI vs. progressive MCI). However, this ignores the continuous nature of time. To capture the temporal dynamics, researchers are shifting toward continuous and ordinal prediction frameworks.
For example, Tsao et al. developed a multi-task learning framework that utilizes high-resolution hippocampal surface morphometry to predict specific cognitive scores up to 48 months in advance (Feature selective temporal prediction of Alzheimer's disease progression using hippocampus surface morphometry). Rather than predicting a static diagnosis, the model maps the continuous decay of structural geometry to cognitive decline.
To explicitly model a patient's position on the timeline, Hagos et al. proposed an interpretable Weighted Siamese Network (Interpretable Weighted Siamese Network to Predict the Time to Onset of Alzheimer's Disease from MRI Images). By reframing the prediction task into an ordinal classification problem, the network takes an input MRI scan and estimates a continuous progression level representing the exact proximity in years to disease onset. This allows clinicians to move away from vague risk percentages and toward concrete temporal windows.
Mapping the Genetic Architecture of Progression
If neuroimaging and biomarkers map the phenotype of disease trajectories, genome-wide association studies (GWAS) seek to identify the genotypic drivers behind them. Traditional imaging genetics often correlates single genetic variants (SNPs) with static brain volumes. Modern trajectory modeling, however, uses the continuous progression metrics themselves as the target phenotype.
Scelsi et al. pioneered this by calculating a multimodal Disease Progression Score (DPS) combining amyloid PET and hippocampal volume, and then running a GWAS directly on this score (Genetic study of multimodal imaging Alzheimer’s disease progression score implicates novel loci). This approach uncovered a novel genetic locus (LCORL) associated with the rate of disease progression, which would have remained hidden using static, single-modality phenotypes.
To understand the causal pathway from genes to cognitive decline, researchers employ mediation analysis. Ding et al. introduced S-GMAS, a genome-wide mediation analysis framework that models how genetic variations influence neurocognitive outcomes through intermediate changes in subcortical brain shapes (S‐GMAS: Genome‐Wide Mediation Analysis With Brain Subcortical Shape Mediators). By treating the 3D geometry of subcortical structures as a continuous mediator, this framework maps the complete causal chain from a specific genetic variant to structural deformation, and ultimately, to cognitive impairment.