AQAL: A Foundation Model Approach to Predicting Neurodiverse Brain Activity

Abstract

Contents

1.Introduction

Autism spectrum conditions affect approximately 1 in 36 children worldwide (CDC, 2023). Understanding how autistic brains process sensory information differently from neurotypical brains is critical for designing inclusive environments, educational strategies, and clinical interventions. However, functional neuroimaging studies are expensive, slow, and inaccessible to most practitioners and families.

Recent advances in brain encoding models — neural networks trained to predict brain activity from sensory stimuli — offer a new paradigm. State-of-the-art multimodal transformers can now predict cortical surface activity from vision, audition, and language inputs with meaningful accuracy across subjects.

We asked: Can a brain encoding model, combined with population-level connectivity data from autism neuroimaging, predict how a neurodiverse brain would respond to arbitrary stimuli?

AQAL is our answer. Rather than requiring individual brain scans, AQAL uses a proprietary statistical transform trained on a large multi-site fMRI corpus to convert neurotypical brain predictions into neurodiverse variants, making neuroscience-informed predictions accessible through a web API.

1.1 Contributions

1. A CPU-based statistical transform that converts neurotypical predictions to neurodiverse predictions using connectivity effect sizes from 871 subjects across 20 clinical sites.
2. A connectivity analysis pipeline identifying 1,002 FDR-corrected significant connections (1,773 uncorrected) between ASD and TD groups across 4,950 tested pairs from 1,545 subjects at 36 sites, with site harmonization and age/sex covariates.
3. A sensory profiling system that maps brain-level divergence onto seven functional networks (visual, auditory, motor, language, social, default mode, salience).
4. A publicly accessible platform (API + web interface) enabling real-time neurodiverse brain prediction from text, audio, or video input.

2.Related Work

2.1 Brain Encoding Models

Brain encoding models learn a mapping from stimulus features to neural activity. Early approaches used linear ridge regression on hand-crafted features (Huth et al., 2016; de Heer et al., 2017). More recent work leverages deep neural network representations, with models like BrainLM (Ortega Caro et al., 2023) and BrainBERT (Thomas et al., 2023) achieving increasingly accurate predictions across modalities.

AQAL builds on this lineage with a proprietary multimodal architecture that accepts vision, audio, and language inputs simultaneously and predicts activity across 20,484 cortical surface vertices using a deep transformer encoder.

2.2 Autism Neuroimaging

Large-scale multi-site autism neuroimaging consortia have provided resting-state fMRI data from over 1,100 individuals, enabling population-level connectivity analysis. Studies using these datasets have consistently identified differences in functional connectivity between ASD and TD groups, particularly in the default mode network, salience network, and limbic regions (Hull et al., 2017; Holiga et al., 2019).

2.3 Gap

No prior work has combined a multimodal brain encoding model with population-level autism connectivity data to generate stimulus-specific neurodiverse brain predictions. AQAL bridges this gap.

3.Methods

3.1 System Architecture

Stimulus (text / audio / video)
        │
        ▼
┌───────────────────────┐
│   Feature Extraction  │  Proprietary multimodal encoders
└──────────┬────────────┘
           │
           ▼
┌───────────────────────┐
│   AQAL Brain Encoder  │  8-layer transformer, 177M params
│   (NT Prediction)     │  → 20,484 vertices × T timesteps
└──────────┬────────────┘
           │
           ▼
┌───────────────────────┐
│   ND Transform        │  Trained on 871 multi-site subjects
│   (ASD Prediction)    │  Per-vertex scale + shift
└──────────┬────────────┘
           │
           ▼
┌───────────────────────┐
│   Analysis &          │  Divergence metrics, sensory profile,
│   Visualization       │  connectivity mapping, interpretation
└───────────────────────┘

3.2 AQAL Brain Encoder

The AQAL brain encoder is a proprietary multimodal transformer trained to predict cortical surface activity from sensory inputs. The model architecture consists of:

• Feature extractors: Dedicated encoders for text (6 layers), video (2 layers), and audio (2 layers) built on state-of-the-art foundation models
• Feature projectors: Per-modality MLPs mapping to a common hidden dimension of 1,152
• Temporal smoothing: Gaussian kernel (size 9) for hemodynamic response alignment
• Transformer encoder: 8 layers, 18 heads, head dimension 64, FF dimension 3,072
• Subject layers: Low-rank output adaptation (bottleneck 2,048) with 0.1 dropout
• Output space: 20,484 vertices on a standard cortical surface mesh (10,242 per hemisphere)

The model operates at 2 Hz temporal resolution. Modality dropout of 0.3 enables predictions from any subset of input modalities.

3.3 Training Dataset

We trained the neurodiverse transform on a large-scale multi-site autism neuroimaging corpus comprising 871 resting-state fMRI scans:

All scans were preprocessed using a standard fMRI preprocessing pipeline with band-pass filtering (0.01–0.1 Hz) and no global signal regression. We extracted mean timeseries from 100 cortical parcels using a validated cortical parcellation, computed Pearson correlation matrices, and applied Fisher's z-transform for variance stabilization, producing 4,950 unique connectivity features per subject.

3.4 Statistical Group Comparison

For each of the 4,950 unique connectivity pairs, we performed independent-samples t-tests with age and sex as covariates and site effects residualized. After Benjamini-Hochberg FDR correction (q < 0.05), 387 of 4,950 connections (7.8%) survived. For reference, 1,065 (21.5%) were significant uncorrected, and 48 (1.0%) survived Bonferroni correction.

The most affected regions ranked by normalized effect size:

These findings are consistent with the autism neuroimaging literature, which identifies limbic hyperconnectivity and default mode network alterations as hallmarks of ASD (Hull et al., 2017).

3.5 Neurodiverse Transform

We derive a proprietary per-vertex transform that converts neurotypical predictions to neurodiverse predictions. For each of 20,484 cortical vertices v:

1. Identify the cortical ROI r(v) to which vertex v belongs.
2. Compute the normalized effect size e_r for that ROI from connectivity t-tests.
3. Compute scale and shift factors using a proprietary mapping function.
4. Apply: &ycirc;_v^ND = &ycirc;_v^NT × scale_v + shift_v

Of 20,484 vertices, 18,067 (88.3%) are affected by the transform, reflecting the widespread but heterogeneous nature of connectivity differences in autism.

3.6 Sensory Profiling

We map vertex-level divergence between NT and ND predictions onto seven canonical functional brain networks:

Scores are normalized to [0, 1] by dividing by the maximum network divergence, providing an interpretable summary of which processing domains are most differently activated.

3.7 Infrastructure

All inference runs on CPU. No GPU is required.

4.Results

4.1 Connectivity Analysis

Analysis of 1,545 subjects from two consortia (36 clinical sites), with site harmonization and age/sex covariates, revealed 1,002 FDR-corrected significant connections out of 4,950 tested (20.2%, q < 0.05). This represents a 2.6x improvement over the single-consortium baseline (387 FDR from 871 subjects). The limbic system showed the largest effect sizes, consistent with known emotional regulation differences. Default mode network showed bilateral temporal alterations. 73 connections survived Bonferroni correction.

4.2 Child-Specific Analysis (Dual-Consortium Combined)

To test whether connectivity differences exist in younger children, we combined two consortia to produce a dedicated child cohort: 548 subjects under age 12 (221 ASD, 327 TD) across 24 sites. Despite 2.5x more subjects than the Consortium I-only child subset, zero connections survived FDR correction (331 uncorrected at p < 0.05, 0 FDR, 0 Bonferroni).

This is a significant negative finding. Resting-state functional connectivity differences between ASD and TD children under 12 are too subtle for group-level statistical detection, even with 548 subjects and FDR correction. The uncorrected count actually decreased from 377 to 331 when adding Consortium II, suggesting some of the Consortium I-only signal was noise. This confirms that the all-ages transform (387 FDR connections) is driven primarily by adolescent and adult data, and that early childhood detection will require either task-evoked fMRI, direct model fine-tuning, or non-fMRI modalities such as EEG or behavioral video analysis.

4.3 Network-Level Differences

4.4 Transform Coverage

The neurodiverse transform affects 18,067 of 20,484 cortical vertices (88.3%), with scale factors ranging from 0.85 to 1.15 and shift factors from \u22120.05 to 0.01. The transform is denser in temporal and limbic regions and sparser in primary motor cortex, matching the known neuroanatomy of autism.

4.5 Real-Time Performance

5.Applications

5.1 Sensory Accessibility Auditing

AQAL can process video of physical spaces (classrooms, offices, retail environments) and generate second-by-second sensory stress predictions for neurodiverse individuals, including flagged high-stress moments and actionable accessibility recommendations.

5.2 Educational Accommodation Design

By profiling how specific educational content differently activates neurodiverse brain networks, AQAL can inform content pacing, optimal break timing, and modality-specific accommodations.

5.3 Clinical Communication

The sensory profile output provides a quantitative, brain-data-grounded framework for communicating how a neurodiverse individual might experience specific stimuli — replacing subjective descriptions with measurable network-level predictions.

6.Limitations & Critique

We present an honest assessment of the structural weaknesses in AQAL v0.1. These are not caveats — they are the core problems that must be solved before AQAL can move beyond an accessibility heuristic toward clinical utility.

6.1 The Transform is Statistics, Not Learning

AQAL multiplies a neurotypical prediction by a fixed map derived from t-tests. This is a linear approximation that discards covariance structure between brain regions. A properly fine-tuned conditional encoder would learn the neurodiverse manifold directly from task fMRI, rather than patching a neurotypical prediction after the fact. GPU fine-tuning was not feasible due to compute constraints, but it is the single most important improvement needed.

6.2 Resting-State to Task-Evoked Leap

The transform extracts connectivity differences from resting-state fMRI (brains doing nothing) and applies them to task-evoked predictions (brains actively processing stimuli). This assumes that resting wiring differences map perfectly to how the brain processes real-time input — a major, imperfect assumption. While resting-state connectivity correlates with task-evoked patterns in the literature, the relationship is not one-to-one, and the degree of error this introduces is unquantified.

6.3 The “Average” Neurodiverse Brain

Autism is inherently heterogeneous. AQAL produces a single, generalized neurodiverse prediction by shifting a neurotypical baseline — this risks stereotyping neurodivergent processing. The same transform is applied to 88.3% of cortical vertices regardless of the individual. A 5-minute individual calibration module is planned but not yet validated.

6.4 Statistical Rigor

The v0.1 release reported 820 uncorrected connections. After implementing FDR correction, site harmonization, age/sex covariates, and expanding to two consortia (1,545 subjects across 36 sites) in v0.3, we identify 1,002 FDR-corrected connections (q < 0.05) — 1,773 uncorrected, 73 Bonferroni. This is a 2.6x improvement over the single-consortium baseline. The core findings in limbic and default mode networks persist and strengthen. Bootstrap confidence intervals (200 iterations) yield a mean 95% CI width of 0.059 across 20,484 vertices. We publish both corrected and uncorrected maps for transparency.

6.5 No Behavioral Ground Truth

There is no validation that predicted ND divergence correlates with actual sensory overload, eye-tracking data, or caregiver report. Without behavioral validation, AQAL is a visualization engine, not a clinical tool. Prospective studies comparing predictions against physiological markers (pupil dilation, galvanic skin response, heart rate variability) are a prerequisite for any clinical claims.

6.6 Dataset Limitations

The training corpus is heterogeneous (20 sites, varying protocols, 86.6% male). Age ranges are mixed without stratification. Results may not generalize to females, young children, or underrepresented populations. Site harmonization and age-stratified transforms are in development.

7.Roadmap

Based on expert review, we divide the improvement plan into two tracks: what can be done on CPU (statistical and validation work) and what requires GPU compute (model retraining).

7.1 Without GPU — Fix the Science

1. FDR-corrected connectivity map. Re-run all 4,950 tests with Benjamini-Hochberg correction, site as random effect, age and sex as covariates. Publish both corrected and uncorrected maps.
2. Age stratification. Split the 871 subjects into developmental bands (child, adolescent, adult) and compute separate transforms per band. No retraining — just regrouping the per-vertex factors.
3. Uncertainty quantification. Bootstrap the transform (200+ iterations) to produce 95% credible intervals for each of 20,484 vertices. Propagate uncertainty to the seven network scores so the UI can display confidence ranges.
4. 5-minute individual calibration. Design a standardized stimulus set, collect brief behavioral or low-cost EEG responses, and fit a per-person scaling vector with ordinary least squares.
5. Behavioral validation. Run prospective studies: child watches a classroom video, AQAL predicts high visual-network divergence, measure pupil dilation, gaze aversion, and caregiver stress rating. Report sensitivity, specificity, and calibration curves.
6. Site harmonization. Standardize preprocessing choices, document protocol differences, release a clean metadata table to reduce noise before any GPU work.
7. Clinical guardrails. Define referral language, risk-flag thresholds, and ethics documentation. Specify what “increased visual and salience divergence” triggers in a pediatric workflow — and what it does not.

7.2 With GPU — Learn the Neurodiverse Brain

1. Conditional encoder fine-tuning. Replace the statistical patch with GPU-accelerated fine-tuning of the 177M-parameter transformer directly on autistic task fMRI, using low-rank adapters per subgroup. Learn the ND manifold instead of multiplying by a fixed map.
2. Developmental models. Train separate encoders for infants and toddlers using prospective datasets. Early detection lives in the first 18 months — adult resting-state maps cannot be extrapolated to babies.
3. Inverse pipeline for screening. Train a brain-to-behavior model: take home video or eye-tracking, predict likelihood of atypical sensory processing using AQAL predictions as a generative prior. A mismatch score across seven networks becomes a risk flag, not a diagnosis.
4. Multi-condition expansion. Joint training for ADHD, sensory processing disorder, and anxiety, plus EEG fusion for portable monitoring.
5. End-to-end uncertainty. Use ensembles or Bayesian neural nets so credible intervals come from the model itself, not post-hoc bootstrapping.

7.3 Toward Early Detection

AQAL does not currently detect autism — it predicts brain activity given a stimulus. Early detection requires the inverse problem: observe naturalistic behavior, infer likely neural processing differences, and flag risk for follow-up. This requires three prerequisites:

1. Train on infants and toddlers. The largest effects in AQAL are limbic and DMN, but early autism markers involve visual attention and social orienting in the first 18 months.
2. Flip the pipeline. Instead of stimulus → brain, build brain → behavior. Compare predicted sensory profiles to real home videos or eye-tracking.
3. Clinical pathway integration. Any risk flag must trigger a structured referral, not a label. Output example: “increased visual and salience divergence relative to age norms, consider M-CHAT follow-up.” Models do not replace clinical assessment.

7.4 Toward Clinical Diagnostic Rigidity (SaMD)

To transition from a heuristic tool to a regulated Software as a Medical Device, AQAL would need:

• ISO 13485 Quality Management System with auditable version control
• Prospective clinical trials comparing predictions against ADOS-2 and ADI-R gold standards
• Physiological validation against real-time markers (HRV, GSR, cortisol)
• FDA De Novo classification pathway with strict Intended Use statement
• Predetermined Change Control Plan (PCCP) for safe algorithm updates
• Proof of clinical utility: clinician + AQAL outperforms clinician alone

8.Conclusion

AQAL is a promising accessibility design tool today. It demonstrates that combining a multimodal brain encoding model with population-level autism connectivity data can produce neuroscience-informed predictions from arbitrary stimuli — without individual brain scans or GPU infrastructure.

However, the current system is a prototype, not a clinical instrument. The statistical transform is an approximation that discards individual variation, the connectivity map requires FDR correction, and no behavioral validation exists. These are not minor caveats — they are the core problems to solve.

The path forward is clear: fix the statistics on CPU first (FDR correction, age stratification, behavioral validation), then invest GPU budget to replace the statistical patch with a learned conditional encoder. Only then can AQAL responsibly explore early detection as a screening aid within standard clinical workflows.

By publishing both the platform and an honest account of its limitations, we aim to accelerate computational neurodiversity research while setting realistic expectations about what the system can and cannot do.

References

Appendix A: Model Parameters

Appendix B: Training Corpus Demographics

Appendix C: System Availability