MMAP: Multimodal Missing-Aware Pretraining for Longitudinal Alzheimer's Prediction
Fiona Kekwick, Matthew Baugh, Bernhard Kainz, Paul M. Matthews, Wenjia Bai
cs.CV · cs.LG
Quick verdict
MMAP pretrains missing-robust image-tabular encoders for AD progression prediction, beating baselines on BACC/AUC; single cohort, no significance tests.
Summary
The paper introduces MMAP, a multimodal pretraining method for 3D brain MRI and tabular clinical data that remains usable when modalities or tabular entries are missing. A 3D ResNet-18 image encoder is pretrained on ADNI with a SigLIP-style sigmoid contrastive loss aligning multimodal projections with image-only and tabular-only projections, plus an MSE reconstruction loss through an image decoder; the tabular encoder is the frozen TabPFNv2 foundation model; cross-modal MLP token generators synthesize features for a missing modality. On two longitudinal ADNI tasks (1,258 subjects; 5,548 scans; pMCI-vs-sMCI conversion and future amyloid status), five-seed results show the best balanced accuracy and AUC among the tested multimodal baselines (HyperFusion, DAFT, TIP, concatenation): amyloid BACC 0.911, AUC 0.940; conversion BACC 0.786, AUC 0.850. However, conversion F1 (0.669) trails TabPFNv2 alone (0.746), and unimodal gains over TabPFNv2 are marginal or absent (e.g., tabular-only conversion AUC 0.820 vs 0.844). Limitations include evaluation on a single cohort, no statistical significance testing, no code availability statement, and small internal inconsistencies between the two Table 2 rows describing the identical full model.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Combining an efficient sigmoid (SigLIP-style) contrastive loss with generative MSE reconstruction improves 3D brain MRI encoder representations, raising conversion BACC from 0.747 (contrastive-only) and 0.768 (reconstruction-only) to 0.786 (Table 2).
- A cross-modal missing-token generator synthesizes features for missing modalities and outperforms learnable and zero tokens (0.786 vs 0.767 I+T BACC), enabling one model to handle missing images, missing tabular data, and incomplete tabular entries.
- On two longitudinal ADNI tasks, MMAP achieves the best BACC and AUC among six tested multimodal configurations (amyloid BACC 0.911/AUC 0.940; conversion BACC 0.786/AUC 0.850) and remains competitive in unimodal settings.
- Optimizing all input scenarios jointly during finetuning (multimodal + unimodal cross-entropies) acts as a regularizer so the model performs well regardless of which modalities are available at test time.
Strengths
Unified handling of all data-availability scenarios in one model
The cross-modal missing-token generator (Section 3) lets a single model serve multimodal, image-only and tabular-only inputs without per-scenario retraining; Table 2 shows generated tokens beat learnable tokens (0.767 I+T BACC) and zero tokens (0.767) at 0.786 BACC for conversion prediction.
Practical combination of foundation models with efficient pretraining
A 3D ResNet-18 encoder pretrained with SigLIP sigmoid loss plus MSE reconstruction (100 epochs on 5,548 scans) plugs into a frozen TabPFNv2 tabular encoder; freezing is justified empirically since finetuning TabPFNv2 degraded performance (Section 4 Implementation).
Fair, seed-replicated baseline comparison
Complementary-loss ablation is clean
Clinically motivated task selection
Weaknesses
Headline claim contradicted by conversion F1
Unimodal advantage over TabPFNv2 is within noise or negative
Single cohort, no significance testing
Internal numeric inconsistency in Table 2
Robustness to incomplete tabular entries is asserted, not measured
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
MMAP (arXiv:2609.26617, cs.CV/cs.LG, MICCAI-format) is a method paper on multimodal pretraining of 3D brain MRI with tabular clinical data under missing-modality and incomplete-tabular conditions, evaluated on ADNI for two longitudinal Alzheimer's tasks: MCI-to-AD conversion (pMCI vs sMCI) predicted 1–3 years ahead, and future amyloid status (PET SUVR > 0.79). The authors claim three things (Section 1): (1) combining a sigmoid contrastive loss with generative reconstruction improves the 3D image encoder; (2) a cross-modal missing-token generator handles missing images, missing tabular data, and incomplete tabular entries; (3) the model matches or beats multimodal (HyperFusion, DAFT, TIP) and unimodal baselines. Claims 1–2 are well supported by ablations; claim 3 holds for BACC/AUC but is contradicted by conversion F1.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is MMAP and what problem does it solve?
MMAP (Multimodal Missing-aware Alignment Pretraining, Imperial College London, arXiv 2026) is an AI method that learns jointly from 3D brain MRI scans and tabular clinical data such as cognitive test scores. Its key feature is robustness to incomplete data: when images are missing, tabular data is missing, or table entries are incomplete, a cross-modal token generator synthesizes features for whatever is absent, so prediction still works with a single model.
How accurate is MMAP at predicting Alzheimer's progression?
On held-out ADNI test data averaged over five random seeds, MMAP reached 0.786 balanced accuracy and 0.850 AUC for predicting whether a patient with mild cognitive impairment would convert to Alzheimer's disease within 1-3 years, and 0.911 balanced accuracy / 0.940 AUC for predicting future amyloid-positive status. Its F1 score for conversion prediction (0.669) was, however, lower than a tabular-only model using TabPFNv2 alone (0.746).
What data was MMAP trained on?
1,258 subjects from the Alzheimer's Disease Neuroimaging Initiative (ADNI), totaling 5,548 T1 MRI scans paired with cognitive and clinical tabular scores (1,442 cognitively normal, 3,370 MCI, 736 AD). 85% of subjects were used for pretraining (including a balanced 2,232-scan subset for finetuning) and 15% were held out for testing in a subject-wise split. Amyloid labels came from PET scans using a standardized uptake value ratio threshold of 0.79.
Is MMAP code publicly available?
The paper does not mention any code, model weights, or data-split release. ADNI itself is publicly available, but access requires registration and a data use agreement, so independent reproduction of the exact numbers would require reimplementation plus an approved ADNI download.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 1 (architecture diagram) is garbled in the text extraction; the architecture was reconstructed from Section 3 prose and caption fragments.
- Possible extraction ambiguity in hyperparameters: λrecons appears as '4' but could be a fraction (e.g., 1/4) in the original PDF.
- The Table 2 internal-inconsistency finding depends on extraction alignment, though the matching I+T BACC (0.786) in both rows suggests correct column alignment.
- Any supplementary material or appendix beyond the extracted main text was not available for review.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.