Abstract
Identifying prediagnostic neurodegenerative disease is a critical issue in neurodegenerative disease research, and Alzheimer’s disease (AD) in particular, to identify populations suitable for preventive and early disease modifying trials. Evidence from genetic studies suggest the neurodegeneration of Alzheimer’s disease measured by brain atrophy starts many years before diagnosis, but it is unclear whether these changes can be detected in sporadic disease. To address this challenge we train a Bayesian machine learning neural network model to generate a neuroimaging phenotype and AD-score representing the probability of AD using structural MRI data in the Alzheimer’s Disease Neuroimaging Cohort (cut-off 0.5, AUC 0.92, PPV 0.90, NPV 0.93). We go on to validate the model in an independent real world dataset of the National Alzheimer’s Coordinating Centre (AUC 0.74, PPV 0.65, NPV 0.80), and demonstrate correlation of the AD-score with cognitive scores in those with an AD-score above 0.5. We then apply the model to a healthy population in the UK Biobank study to identify a cohort at risk for Alzheimer’s disease. This cohort have a cognitive profile in keeping with Alzheimer’s disease, with strong evidence for poorer fluid intelligence, and with some evidence of poorer performance on tests of numeric memory, reaction time, working memory and prospective memory. We found some evidence in the AD-score positive cohort for modifiable risk factors of hypertension and smoking. This approach demonstrates the feasibility of using AI methods to identify a potentially prediagnostic population at high risk for developing sporadic Alzheimer’s disease.
Introduction
The primary challenge in Alzheimer’s research is identifying the disease in its pre-diagnostic stage—years or even decades before clinical symptoms emerge. While genetic studies indicate that brain atrophy begins long before a clinical diagnosis, detecting these subtle, sporadic changes in the general population remains difficult. Current gold-standard biomarkers, such as PET scans for amyloid or tau, are expensive, invasive, and not widely accessible. This study explores whether structural MRI—a common, non-invasive imaging tool—combined with advanced Bayesian machine learning can serve as a scalable screening tool to identify at-risk populations for early intervention.
Methods
The researchers implemented a Bayesian Neural Network (BNN) to move beyond deterministic classifications and incorporate uncertainty estimation.
Feature Extraction and Preprocessing
Using FreeSurfer, 155 features were extracted from structural MRI scans, including regional cortical thickness and volumes (e.g., hippocampus, amygdala, thalamus). To ensure the model learned disease-specific patterns rather than general aging effects, confounds such as age, sex, and total intracranial volume were regressed out using linear models fitted on the training set.
The Bayesian Approach
Unlike standard neural networks, this model utilizes Monte Carlo (MC) dropout. By executing 50 forward passes for every scan with different random "dropout masks," the model generates a distribution of results. The mean of this distribution serves as the "AD-score" (probability of disease), while the standard deviation represents the model's uncertainty.
Datasets
The model was trained and internally tested on the ADNI (Alzheimer’s Disease Neuroimaging Initiative) dataset (n=736), a well-characterized research cohort. It was then validated on the NACC (National Alzheimer’s Coordinating Centre) dataset (n=5,209), which reflects more heterogeneous "real-world" clinical practice. Finally, the model was applied to the UK Biobank (n=37,104), a target population of individuals with no dementia diagnosis at the time of imaging.
Results
The model demonstrated high performance, achieving an AUC of 0.97 and 92% accuracy on the ADNI test set. On the noisier NACC dataset, it maintained a respectable AUC of 0.79 (74% accuracy).
Uncertainty as a Quality Filter
A key finding was the relationship between uncertainty and accuracy. The model exhibited significantly higher uncertainty when making incorrect predictions. This allows the BNN to act as a quality filter: in a clinical setting, high uncertainty signals that the AI is unsure, prompting a manual review rather than relying on a potentially false result.
Identifying the "Silent" Phenotype in UK Biobank
When applied to the "healthy" UK Biobank cohort, the model flagged 1,304 individuals (3.4%) as AD-score positive, indicating an AD-like neuroimaging phenotype despite the absence of a clinical diagnosis.
Cognitive and Risk Factor Validation
Validation against cognitive data confirmed that the "at-risk" label has functional relevance. The AD-positive group in the UK Biobank performed significantly worse on tests of fluid intelligence (Effect size -0.35, CI -0.46 to -0.21), numeric memory, and reaction time. Furthermore, this group showed higher rates of hypertension and smoking history, reinforcing the link between modifiable cardiovascular risk factors and neurodegeneration.
Discussion
This study demonstrates the feasibility of using AI to identify potentially prediagnostic populations at high risk for sporadic Alzheimer’s disease. The Bayesian approach consistently outperformed non-stochastic versions of the same architecture by providing a measure of confidence in its predictions.
However, limitations remain. The UK Biobank cohort may suffer from selection bias toward healthier participants. Additionally, longitudinal follow-up is required to confirm how many AD-score positive individuals eventually progress to a clinical diagnosis. Future work should focus on integrating this MRI-based AI score with other biomarkers, such as polygenic risk scores or blood-based markers, to further refine early detection strategies.
Summary updated