🧠 neurology

Quantifying Uncertainty in Alzheimer's Disease Progression Modelling: A Variational Disease Progression Score Framework

This paper introduces a variational Disease Progression Score framework that leverages multimodal biomarkers and the amyloid cascade hypothesis to generate biologically interpretable, uncertainty-quantified individual prognoses for Alzheimer's disease progression, demonstrating high accuracy and dynamic adaptability on the ADNI cohort.

Original authors: Ngamsaowaros, T., Bodala, I., Michopoulou, S., Niranjan, M.

Published 2026-08-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ngamsaowaros, T., Bodala, I., Michopoulou, S., Niranjan, M.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Alzheimer's disease is a slow, relentless thief that steals memory and identity, but it does not steal from everyone in the same way or at the same speed. For decades, doctors have understood that the disease follows a biological sequence: first, toxic proteins called amyloid and tau begin to accumulate in the brain; then, brain tissue begins to shrink; and finally, thinking and daily functioning decline. However, this sequence plays out differently for every person. Some people show signs of the disease in their sixties, while others do not show symptoms until their eighties. Some progress rapidly, while others decline slowly. This variation makes it incredibly difficult to predict how a specific individual will fare, especially when the medical data available is often sparse, collected at irregular times, and missing key pieces. Without a clear picture of where a person stands on this invisible timeline, it is hard to know when to intervene or how to design clinical trials that might actually help.

A team of researchers at the University of Southampton has developed a new way to map this invisible journey. They created a computer model that acts like a universal ruler for the disease, placing every patient on a single, continuous timeline based on their unique biology. Instead of guessing a patient's stage by looking at a single test result, the model weaves together age, genetics, education, and various brain scans and blood tests to estimate two crucial things: when the disease likely started for that person and how fast it is moving. Crucially, the model does not just give a single answer; it also calculates how confident it is in that answer, providing a range of possibilities rather than a fixed prediction. This approach allows doctors to see not just where a patient is today, but to project their future path with a clear understanding of the uncertainty involved.

The researchers tested their system using data from the Alzheimer's Disease Neuroimaging Initiative, a massive public database containing information from hundreds of people. They fed the model basic information available at a patient's first visit, such as their age, sex, years of education, and a specific genetic marker known to influence risk. From this starting point, the model generated a personalized disease score for each person. Remarkably, even though the model was never told the patients' diagnoses, it successfully sorted them into the correct groups. People who were cognitively normal clustered at one end of the timeline, those with mild impairment sat in the middle, and those with Alzheimer's disease clustered at the other end. The model separated these groups with a high degree of accuracy, proving that it had learned the underlying biological rhythm of the disease without being explicitly taught the labels.

Beyond sorting patients, the model reconstructed the order in which different parts of the brain are affected. It confirmed the established scientific view that the accumulation of amyloid protein happens first, followed by the buildup of tau protein, which then triggers the shrinking of brain regions responsible for memory, and finally leads to cognitive decline. The model even quantified how strongly one stage drives the next. It found that the link between tau protein and brain shrinkage was the strongest connection, while the link between amyloid and tau was weaker but still consistent. This suggests that while amyloid may be the spark that starts the fire, the spread of tau is what causes the most significant damage to brain tissue.

One of the most powerful features of this framework is its ability to improve over time. In the real world, patients return for follow-up visits, providing new data points. The researchers showed that as new measurements came in, the model could update its prediction for that specific person. It could refine its estimate of how fast the disease was progressing and narrow down the range of uncertainty. For example, if a patient's initial blood test suggested a slow progression, but their follow-up brain scan showed rapid shrinkage, the model would adjust its forecast to reflect this faster decline. This dynamic updating mirrors how a doctor revises a diagnosis as a patient's condition changes, moving from a broad population-based guess to a highly specific, personalized prognosis.

The study also revealed how different factors influence the speed and timing of the disease. The researchers found that carrying two copies of a specific genetic variant, known as APOEε4, was the strongest predictor of an earlier disease onset. People with this genetic profile were estimated to start their disease journey years earlier than those without it. Education also played a role, with higher levels of education associated with a later onset of symptoms, supporting the idea that a "cognitive reserve" can delay the appearance of problems even if the underlying disease is present. Interestingly, the model showed that the rate of decline was not the same for everyone; it tended to be fastest in people in their mid-to-late seventies, while progression in the very elderly was more variable and often slower.

While the model performed well, the researchers were careful to note its limitations. The predictions were most accurate for cognitive tests and brain volume measurements, but less precise for certain fluid biomarkers, which are harder to measure consistently. The model also relies on a simplified view of the disease, assuming a single path of progression, which may not capture every subtype of Alzheimer's. Furthermore, while the model can predict the future based on current data, it has not yet been tested on completely different populations outside the study group. Despite these caveats, the work represents a significant step forward. It moves the field away from static snapshots of disease toward a dynamic, probabilistic understanding that acknowledges uncertainty. By providing a tool that can track an individual's unique path and update its predictions as new information arrives, this framework offers a more realistic and hopeful foundation for managing Alzheimer's disease in the future.

Technical Summary: Quantifying Uncertainty in Alzheimer's Disease Progression Modelling

Problem Statement
Predicting the course of Alzheimer's disease (AD) for individual patients is hindered by the heterogeneity of disease expression and the sparsity of longitudinal clinical data. Existing methods face a trade-off: classical Disease Progression Score (DPS) models lack inter-biomarker relationship modeling; Ordinary Differential Equation (ODE) models require dense sampling rarely available in real-world cohorts; Mixed-effects models often rely on restrictive trajectory assumptions and lack predictive uncertainty; and deep learning approaches, while flexible, operate as "black boxes" with limited interpretability and uncertainty quantification. Consequently, no existing method simultaneously delivers biological interpretability, calibrated uncertainty quantification, and reliable generalization to previously unseen patients.

Methodology
The authors propose a Variational Disease Progression Score (DPS) framework that integrates baseline demographic/clinical features with sparse, asynchronous, multimodal longitudinal data. The framework consists of two primary components:

Variational Encoder for Subject-Specific Parameters:
- Baseline features (age, sex, education, APOEε4 genotype, and baseline Clinical Dementia Rating-Sum of Boxes [CDR-SB]) are encoded via a neural network into posterior probability distributions over two latent subject-specific parameters: progression rate ( $\alpha_i$ ) and onset time ( $\beta_i$ ).
- These parameters define a continuous Disease Progression Score (DPS) that time-warps chronological age onto a shared latent disease timeline: $s_i(t) = \alpha_i \frac{(t-75)}{10} + \beta_i$ .
- Uncertainty is quantified by treating $\alpha_i$ and $\beta_i$ as probabilistic variables (Log-normal and Gaussian, respectively) rather than point estimates, learned via variational inference.
Cascade Biomarker Trajectory Model:
- Biomarker trajectories are modeled as a cascade of logistic functions structured according to the amyloid cascade hypothesis. The topology is fixed: Amyloid-β ( $A\beta$ ) $\rightarrow$ Phosphorylated Tau ( $p$ -tau) $\rightarrow$ Neurodegeneration (Hippocampus and Entorhinal cortex volumes) $\rightarrow$ Cognitive Decline (ADAS-13 and CDR-SB).
- Each biomarker trajectory is a function of the DPS and its upstream "parent" biomarkers. The influence of parent biomarkers is modeled as the integral of their logistic trajectories, capturing the cumulative nature of neuropathological damage.
- The model distinguishes between epistemic uncertainty (model uncertainty regarding subject parameters, estimated via Monte Carlo sampling) and aleatoric uncertainty (inherent measurement noise, modeled as homoscedastic for fluid/imaging data and heteroscedastic for cognitive scores).

Training and Inference

Training: The framework uses an Expectation-Maximization (EM)-like algorithm alternating between an E-step (updating the encoder to infer subject posteriors) and an M-step (updating population-level cascade parameters and noise variances).
Sparse Data Handling: Loss functions are computed only over available biomarkers at each visit, allowing the model to utilize incomplete longitudinal records.
Dynamic Personalization: For new longitudinal observations, the framework employs a sequential Bayesian updating scheme. The posterior inferred from baseline data serves as the prior for subsequent visits, allowing the model to refine individual trajectories and reduce uncertainty as data accrues.

Key Results
Evaluated on the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort ( $N=819$ ), the framework demonstrated:

Unsupervised Clinical Alignment: Without using diagnostic labels as inputs, the inferred DPS successfully separated Cognitively Normal (CN), Mild Cognitive Impairment (MCI), and Alzheimer's Disease (AD) groups. The Area Under the Curve (AUC) for CN vs. AD separation was 0.98.
Biological Plausibility: The model recovered biomarker ordering consistent with established pathology: $A\beta$ accumulation preceded $p$ -tau, which preceded neurodegeneration and cognitive decline. The estimated cascade strengths showed the strongest links between tau pathology and medial temporal atrophy, and the weakest (yet most reproducible) link between $A\beta$ and $p$ -tau.
Subject-Specific Generalization: The model successfully inferred progression parameters for entirely unseen subjects (validation cohort) using only baseline features. Cognitive measures (ADAS-13, CDR-SB) showed the strongest predictive alignment ( $r=0.61$ and $0.71$, respectively).
Uncertainty Calibration:
- 95% credible intervals achieved 89–98% empirical coverage across biomarkers in the unseen validation cohort.
- Dynamic fine-tuning reduced parameter uncertainty significantly (e.g., a 66.4% reduction in progression rate uncertainty for a representative subject) and improved point prediction accuracy for most biomarkers, particularly cognitive scores.
- Calibration was generally well-maintained, though it degraded slightly for $p$ -tau and CDR-SB after fine-tuning, likely due to the dominance of frequent cognitive sampling over sparse fluid biomarkers.

Significance and Claims
The paper claims to provide a biologically interpretable, uncertainty-aware index of disease severity that bridges the gap between population-level inference and personalized prognosis. Key contributions include:

Integration of Interpretability and Uncertainty: Unlike deep learning "black boxes," this framework explicitly models the amyloid cascade hypothesis while providing calibrated probabilistic outputs.
Generalization to Unseen Subjects: The variational encoder allows the model to generate individualized prognoses for new patients from baseline data alone, without re-fitting the entire model.
Dynamic Clinical Utility: The framework supports the clinical workflow of refining prognoses over time as new longitudinal data becomes available, dynamically updating both the predicted trajectory and its associated uncertainty.

The authors conclude that while the framework offers a probabilistic foundation for precision monitoring, its reliability on out-of-distribution datasets and its capacity for long-term extrapolation beyond observed follow-up windows require further external validation.

Technical Summary: Quantifying Uncertainty in Alzheimer's Disease Progression Modelling

More like this