NormMoX: A Normative Mixture of Experts for Explainable Autism Classification from Structural and Functional MRI

Muhammad Hasan Masrur, Fani DeligianniarXiv 2026 · arXiv

This paper introduces NormMoX, a novel framework for classifying Autism Spectrum Disorder (ASD) using brain imaging. By shifting the focus from raw data to deviations from a "typical brain," the authors provide a model that is not only more accurate but also significantly more explainable for clinicians and researchers.

TL;DR

NormMoX uses "normative modeling" to classify ASD by comparing individual brain scans against a lifespan-defined "typical brain" reference. By fusing structural and functional MRI through a specialized mixture-of-experts architecture, it achieves superior accuracy (0.728 AUROC) and provides biologically grounded explanations for every prediction.

What Problem Does This Paper Address?

Classifying ASD from MRI scans is notoriously difficult due to the condition's high heterogeneity—brain differences vary wildly from one person to the next. Furthermore, standard machine learning models often accidentally learn to identify age, sex, or scanner-specific noise rather than the actual signatures of ASD.

Existing multimodal models (combining structural and functional data) also suffer from two major gaps:

  1. The "Black Box" Fusion: They often fail to show which modality (e.g., brain shape vs. brain connectivity) is driving a specific diagnosis.
  2. Arbitrary Explanations: Current explainable AI (XAI) methods often explain decisions relative to a random baseline rather than a biologically meaningful "typical" state.

Methods

The core innovation of NormMoX is the use of a shared normative reference frame.

  1. Normative Modeling: Instead of raw values, the model uses "deviations" (z-scores). The authors used large external datasets (over 20,000 healthy individuals) to build "growth charts" for the brain. A participant's scan is mapped to these charts to see how much their brain measures deviate from what is expected for their age and sex.
  2. Normative Experts: The system uses two dedicated ensembles of neural networks—one for structural MRI (sMRI) and one for functional MRI (fMRI). The sMRI expert is specifically designed to look at the magnitude of deviations ($|z|$), capturing the fact that ASD can involve both abnormally thick and abnormally thin cortical regions.
  3. MoX Fusion Head: A specialized fusion layer combines the scores from both experts. Crucially, it calibrates these scores against typical controls, allowing it to handle missing data (e.g., if a patient only has an sMRI scan) without needing complex data imputation.
  4. Typical-Brain Attribution: To explain a prediction, the model uses Integrated Gradients anchored to $z=0$ (the mathematical representation of a perfectly "typical" brain). This tells the user exactly which brain regions moved the needle toward an ASD diagnosis.
Fig. 1: Overview of NormMoX. Normative models turn sMRI and fMRI measures into deviations from typical brain. One expert per modality scores the deviations, MoX fuses the scores relative to typical controls, and integrated gradients from the typical brain explain predictions.

Datasets

  • ABIDE I: The primary dataset for training and evaluation, consisting of 1,035 participants (505 ASD, 530 controls) across 20 different imaging sites.
  • External Reference Sets: Lifespan normative models were pretrained on massive healthy cohorts (approx. 23,000 for sMRI and 22,000 for fMRI) to define the "typical" brain baseline.

Key Insights and Ideas

  • Deviations are Better than Raw Data: Training on normative deviations outperformed training on raw values or data normalized within a single study. This suggests that the "lifespan" context provided by external big data is vital for finding ASD signals.
  • ASD is Often Unimodal: The study found that many participants showed strong evidence for ASD in only one modality (either sMRI or fMRI). A model that requires both or averages them poorly would miss these cases.
  • Explainability as a Sanity Check: By anchoring explanations to the typical brain, the authors could confirm that their model was picking up on known ASD markers, such as hypoconnectivity in limbic networks and enlarged ventricles.
Fig. 2: Mean contribution to the fused prediction in autistic test par- ticipants, where consistent across folds (|mean| ≥2 SD). Red: to- wards ASD; blue: towards control. (a) Ribbon width: contribution of each network pair. (b) sMRI: cortical thickness on inflated surfaces.

Key Results

  • Classification Performance: NormMoX achieved an AUROC of 0.728, significantly outperforming linear baselines, single-modality experts, and standard deep fusion models.
  • Generalization: When tested on entirely unseen imaging sites (Leave-One-Site-Out), the model maintained a robust AUROC of 0.711.
  • Clinical Utility: The model identified that roughly 29% of autistic participants were "flagged" by both modalities, while many others were only detectable in one, highlighting the necessity of the multimodal approach.
Fig. 3: (a) Mean z (95% CI) of the most separating features; open circles: no significant mean shift (contribution through |z|). (b) Par- ticipants flagged in fMRI only, sMRI only or both (10-repeat mean).

Limitations and Future Work

  • Site Calibration: While the lifespan normative models are fixed, the system still requires a small set of local controls from any new imaging site to calibrate the baseline.
  • Scope: The current study focused on ABIDE I. The authors suggest that this framework could be easily extended to other neurodevelopmental conditions like ADHD or even neurodegenerative diseases.
  • Data Requirements: The model relies on high-quality structural and functional preprocessing, which can be computationally intensive.

The implementation is open-source and available at https://github.com/Novelex/NormMox.

Summary updated

Abstract

While structural and functional magnetic resonance imaging (sMRI, fMRI) offer profound potential for researching autism spectrum disorder (ASD) and understanding individual neurodivergence, current classifiers lack the explanatory power needed for scientific study, clinical use and personal-level insights. Most models learn from imaging features that also vary with age, sex, scanner site and head motion. They fuse the two modalities without a common reference and explain decisions against arbitrary reference inputs rather than a typical brain. We present NormMoX, a normative mixture of experts that makes the typical brain the shared reference for prediction, fusion and explanation. Normative models trained on large healthy cohorts give the value each brain measure is expected to take for a person's age and sex. NormMoX learns from each person's deviations from these expected values. For each modality, a neural network expert turns the deviations into a score of evidence for ASD. A fusion head, MoX, compares each score with those of typical controls before combining them. It therefore needs no imputation when a modality is missing and shows in which modality each person is atypical. On 1,035 participants of ABIDE I, NormMoX reached an AUROC of $0.728 \pm 0.004$ in repeated cross-validation, above linear, single-modality and deep fusion baselines. NormMoX thus shows not only whether a person is likely to have ASD, but where, and in which modality, their brain departs from typical. The implementation is available at https://github.com/Novelex/NormMox.
BibTeX
@misc{masrur2026normmox,
  author = {Muhammad Hasan Masrur and Fani Deligianni},
  title = {{NormMoX: A Normative Mixture of Experts for Explainable Autism Classification from Structural and Functional MRI}},
  year = {2026},
  eprint = {2611.03349},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.03349},
}
Figure