A cheap blood test can now flag many people with mild cognitive impairment who are heading toward Alzheimer's disease. In one ADNI cohort it still leaves about a third of patients in an ambiguous middle zone. Brain MRI carries information that could resolve some of those cases, because it shows where the brain is shrinking. But MRI is expensive, scanner-dependent and often unavailable where screening is needed. This note proposes using MRI only during training. A model that takes blood measurements alone (a plasma protein panel plus metabolites, the small molecules circulating in blood) is trained to predict two things: a patient's future conversion to dementia, and that patient's three-dimensional atrophy pattern. At deployment no scan is required. The MRI works as "privileged information": it is available to the learner and withheld from the test-time model. The bet is that a dense anatomical target constrains a model trained on roughly 200 patients far better than a single binary outcome label does.
Two practical obstacles stand in the way, and each maps onto one of the research directions. First, metabolite measurements come in arbitrary, platform-specific units. We therefore represent them as within-sample ranks, following MIRTH, so that the blood-to-anatomy mapping can be learned in one large cohort and reused in another. Second, the clinical cohort where the outcome is known is tiny. We therefore learn the mapping from metabolites to atrophy in UK Biobank, which has both brain MRI and a 327-metabolite NMR panel, and only adapt it on the small clinical cohort.
Core Claim
The claim is that, at fixed deployment inputs (blood only), a predictor trained to also reconstruct a volumetric atrophy representation will beat the same predictor trained on outcome labels alone. Specifically, it will have better discrimination, better calibration, and fewer patients left in the indeterminate risk zone. The gain should be largest when the clinical training set is smallest.
Three findings motivate this. Structural MRI loses sensitivity early in the disease: in Residual and Plain 3D CNNs, which trained 3D networks on ADNI scans, the AUC for early MCI versus controls is 0.57. Plasma p-Tau217 is already informative at this stage: Predicting Progression from MCI using Survival Analysis reports a 4-year AUROC of 0.787, yet its dual-threshold rule leaves 31.4% of patients unclassified. Finally, Multi-organ MRI aging clocks shows by association studies in about 300,000 people that plasma metabolites track brain-MRI age gaps, so a blood-to-anatomy signal exists at scale.
The closest method-level precedent is Transferable Lesion-Supervised Speech Representations. That work trains an accessible modality, speech, against 3D lesion maps in 639 stroke patients and transfers the result to a sub-cohort of 110. We borrow its two-stage logic. We differ in three ways: the accessible modality is molecular, the disease is neurodegenerative, and the endpoint is time-to-conversion rather than lesion geometry.
Research Program
Stage one builds the anatomical teacher, which is the volumetric target the blood model learns to reproduce. We compare three teacher targets, because the choice is the main open question. The first is a scalar brain-age gap in the style of the multi-organ aging-clock paper. The second is a vector of regional volumes, harmonized across scanners with Neuroharmony from A Machine Learning Model to Harmonize Volumetric Brain MRI. That method removed scanner bias across 43 scanners while preserving disease signal (ΔAUC −0.006). The third is a learned whole-volume embedding: the 128-dimensional contrastive representations of Learning heritable multimodal brain representation, which were trained on UK Biobank T1/T2 pairs and lowered brain-age error to 3.57 years.
Stage two learns the student, a small network that maps rank-transformed metabolites to the teacher target, in UK Biobank. It is fit with a soft-margin distance loss borrowed from Cross-Modal Distillation For Widely Differing Modalities. The margin matters because that paper showed that forcing exact feature matching between very different modalities makes the student memorize modality-specific noise. A blood profile cannot reproduce an MRI embedding exactly, and it should not be penalized for failing to.
Stage three happens in the small ADNI MCI cohort. A survival head is trained on the student's predicted atrophy representation together with the accessible plasma panel (p-Tau217, Aβ42/40, NfL, GFAP). During this stage the subject's own MRI-derived target is kept as an auxiliary loss on the training split only.
Metabolite inputs are rank-transformed so that the stage-two mapping does not depend on absolute units. We take seriously the warning in the research file's harmonization discussion that rank transforms without conditioning on diagnosis can scramble disease signal in small cohorts. For that reason we also test ranks computed within diagnostic strata of the reference cohort.
One data caveat must be stated up front. The file documents ADNI plasma protein biomarkers but does not document an ADNI metabolomic panel compatible with UK Biobank's NMR features. If no such overlap exists, stage three falls back to the plasma protein panel as the student input, with the MRI auxiliary loss intact. The metabolite-transfer arm would then be evaluated only within UK Biobank, using incident dementia diagnoses from ICD-10 codes.
Baselines
Naive reference. The do-nothing reference is the p-Tau217-only Random Survival Forest from Predicting Progression from MCI using Survival Analysis. That paper reports a 4-year conversion AUROC of 0.787 and 68.6% decisive stratification on 201 ADNI MCI subjects. We reproduce it on identical subjects and folds. It is the strongest accessible competitor because it is the clinical default and it already outperforms amyloid PET (0.758) in that paper.
Strongest blood-only learned model. This is the same survival model given the full plasma panel, plus metabolites where available, with no MRI supervision. Its hyperparameter search budget equals the proposal's.
MRI-at-test-time upper reference. This is the stability-selected sparse model of Sparse learning and stability selection, which reports a 48-month AUC of 0.8587 from MRI, cognitive and APOE features. It uses inputs we deliberately withhold, so the question is how much of the gap between 0.787 and 0.8587 a blood-only student can close.
Isolating ablation. The key ablation runs the identical architecture with the auxiliary atrophy loss removed. A second ablation replaces rank inputs with per-cohort z-scores.
What counts as beating the baselines. A win requires two things. The first is a gain in time-dependent AUROC over the p-Tau217 baseline whose bootstrap 95% interval excludes zero. The second is a reduction in the indeterminate fraction below 31.4% at matched 90% sensitivity and specificity cutoffs. Both must hold under five repeats of nested cross-validation.
Why not just…?
Why not just use p-Tau217 alone? This is a strong threat, because a single assay is cheap and well validated (DOR 42 in Blood biomarkers in MCI conversion). However, that meta-analysis also reports weaker performance in non-amnestic MCI and large variation across assay platforms. The proposal targets exactly the indeterminate 31.4% zone. If the gain appears only in cases p-Tau217 already decides, the idea fails.
Why not just predict regional MRI volumes from blood and plug them in as features? This two-stage imputation is the most likely equivalent. It is the regional-volume teacher arm without the joint auxiliary loss in stage three, and we run it as a separate condition. If it matches the joint model, we will report imputation as the method and drop the joint-training claim.
Why not skip native 3D representations and use regional summaries? End-To-End AD Diagnosis found that smaller 3D networks beat larger ones on ADNI, and 3D CNNs overfit small cohorts. This is why no 3D network is ever trained on the clinical cohort: the volumetric embedding is learned in UK Biobank only. Whether it adds anything beyond regional volumes is settled empirically by the comparison of teacher targets.
Why not just z-score metabolites instead of ranking? Z-scoring keeps absolute scale, which the clinical blood-biomarker cutoffs depend on. The z-score ablation shows whether ranks are needed for transfer from UK Biobank to ADNI.
Proposed Experiments
Primary experiment. We train on ADNI MCI subsets of 50, 100 and all roughly 200 subjects. Metrics are 4-year time-dependent AUROC, C-index, calibration slope, and the indeterminate fraction. The decisive check is a learning curve: the auxiliary-loss gain should shrink as the sample size grows. A flat gain would suggest that the loss merely adds features rather than regularizing a small-sample model.
Stage-two fidelity check. In held-out UK Biobank participants, we measure how well blood predicts each teacher target, reported as R² per region. The ceiling is likely to be low: CSF metabolomics reports a mean R² of 0.024 for genotype-to-metabolite prediction. If hippocampal and ventricular targets fall near zero, the premise fails early and cheaply.
Transportability check. We report results stratified by amnestic versus non-amnestic MCI. HABS-HD, described in Assessing the Generalizability of ADRDs Subtypes, lets us test whether the MRI teacher's harmonized regional targets preserve atrophy subtypes across Black, Hispanic and non-Hispanic white participants. The file documents no blood data for HABS-HD, so no fairness claim about the blood student will be made from it.
References
- MIRTH: Metabolite Imputation via Rank-Transformation and Harmonization
- Residual and plain convolutional neural networks for 3D brain MRI classification
- Predicting Progression from Mild Cognitive Impairment to Alzheimer’s Disease using Survival Analysis
- Multi-organ MRI digitizes biological aging clocks across proteomics, metabolomics, and genetics
- Transferable Lesion-Supervised Speech Representations for Post-Stroke Modelling
- A Machine Learning Model to Harmonize Volumetric Brain MRI Data for Quantitative Neuroradiologic Assessment of Alzheimer Disease
- Learning heritable multimodal brain representation via contrastive learning
- Cross-Modal Distillation For Widely Differing Modalities
- Sparse learning and stability selection for predicting MCI to AD conversion using baseline ADNI data
- Blood biomarkers in MCI conversion to Alzheimer’s disease: a systematic review and meta-analysis
- End-To-End Alzheimer's Disease Diagnosis and Biomarker Identification
- Cerebrospinal fluid metabolomics identifies 19 brain-related phenotype associations
- Assessing the Generalizability of ADRDs Subtypes Across Diverse Cohorts HABS‐HD and ADNI