Abstract
Normative modelling is an emerging method for quantifying how individuals deviate from the healthy populational pattern. In this study, we assessed normative models based on deep autoencoders using structural neuroimaging data from patients with Alzheimer's disease (n = 206) and mild cognitive impairment (n = 354). We first trained the autoencoder on an independent dataset (UK Biobank dataset) with 11,034 healthy controls. Then, we estimated how each patient deviated from this norm and established which brain regions were associated to this deviation. As expected, we found that patients exhibited deviations according to the severity of their clinical condition. The model identified medial temporal regions, including the hippocampus, and the ventricular system as critical regions for the calculation of the deviation score. Overall, the normative model had comparable cross-cohort generalizability to traditional classifiers.
Introduction: The Challenge of Heterogeneity
Diagnosing neurodegenerative diseases like Alzheimer’s (AD) is complicated by the overlap between pathological structural changes and healthy aging. Traditional machine learning approaches often utilize "case-control" designs, which require large, balanced datasets and frequently struggle with cross-cohort generalizability.
This study explores normative modelling, an anomaly-detection approach that learns the range of "normal" brain variation first. Instead of binary classification, the model quantifies how much an individual's brain differs from a healthy reference of the same age and sex.
Methodology: Adversarial Autoencoders
The researchers employed an Adversarial Autoencoder (AAE) to define the biological norm. The architecture consists of an encoder that compresses 101 regional brain volumes into a low-dimensional latent space, and a decoder that reconstructs the original data.
Crucially, the model is conditioned on age and sex, allowing it to predict what a healthy brain should look like for a specific demographic. The "deviation score" is calculated as the Mean Squared Error (MSE) between the actual scan and the model’s reconstruction. A high error indicates a pathological deviation from the learned healthy norm.
Multi-Cohort Validation and Results
The model was trained on 11,034 healthy controls from the UK Biobank and evaluated across five independent clinical datasets: ADNI, AIBL, ARWiBo, OASIS-1, and MIRIAD.
Disease as a Spectrum
A key finding was that deviation scores accurately reflected clinical severity. Rather than a simple binary flag, the scores followed a clear trajectory: Healthy Controls (HC) < Mild Cognitive Impairment (MCI) < Alzheimer’s Disease (AD).
Performance Benchmarks
The normative model achieved strong discriminative power, with AUC scores for AD vs. HC ranging from 0.84 to 0.98. For MCI detection, AUCs ranged between 0.60 and 0.72. When compared to traditional classifiers like Relevance Vector Machines, the normative model demonstrated superior consistency and generalizability when applied to entirely new populations, despite never seeing a patient scan during training.
Anatomical Insights and Future Directions
By analyzing reconstruction errors per region, the model identified the medial temporal lobe (including the hippocampus) and the ventricular system as primary drivers of the deviation score—aligning perfectly with established clinical knowledge of AD-related atrophy.
While promising, future iterations could benefit from explicit harmonization of inter-scanner variability and the application of longitudinal data to track individual deviation trajectories over time. Moving toward voxel-wise analysis may also reveal more granular patterns of neurodegeneration.
Summary updated