An Enhanced Conditional Variational Autoencoder-Based Normative Model for Neuroimaging Analysis
Normative modelling in neuroimaging provides a powerful framework for quantifying individual deviations from expected brain measures as a function of relevant covariates. While most established approaches have focused on analysing distinct variables in isolation, there is a growing need for deep learning-based methods capable of handling multiple response variables simultaneously. Conditional variational autoencoders (cVAEs) have previously been applied in this context; however, existing inference methods still face challenges in providing reliable probabilistic predictions. In this study, we introduce an enhanced cVAE-based framework for normative modelling of neuroimaging data. Our key contribution is the development of a novel inference method that integrates latent space sampling and bootstrapping. This method directly leverages the generative nature of cVAEs to estimate the conditional distributions of brain measures. We demonstrate the effectiveness of this approach using white matter hyperintensity (WMH) volumes as a test case for developing, testing, and validating the model. Our dataset includes 8,551 normotensive and 18,180 hypertensive participants from the UK Biobank. We evaluated our method against three well-established and popular normative modelling techniques, including Generalised Additive Models for Location, Scale, and Shape (GAMLSS), Multivariate Fractional Polynomial Regression (MFPR), and Hierarchical Bayesian Regression (HBR). Our results indicate that the proposed cVAE-based framework achieves comparable performance across various metrics, while capturing individual deviations that correlate with the severity of hypertension. The model is particularly well-suited for high-dimensional, non-linear data, making it a robust tool for assessing individual deviations in brain structure. This enhanced normative modelling architecture paves the way for more nuanced and reliable assessments of brain health, with potential implications for personalised medicine and early detection of neurological disorders.