A cross-sectional brain-imaging study can tell you that two things travel together — say, depression liability and degraded white matter — but not which one drives the other. Shen et al. needed genetic instruments, not correlations, to show that white matter damage is a consequence of depression liability rather than a precondition. Meanwhile, the autoencoder community has built exactly the machinery one would want for this problem — models that compress high-dimensional scans into a few latent numbers — but the most careful practitioners admit the core ambiguity is unresolved: Kumar et al. compress three Alzheimer's imaging modalities into a Disease Severity Index, yet cannot tell whether patients who look different are on different causal pathways or merely at different points on one pathway. Causal autoencoders exist (VACA, CausalVAE), but only on toy graphs and celebrity faces, never on volumetric medical imaging.
This proposal argues that the obstacle is not just model architecture but data hygiene and scale, and that two seemingly unrelated literatures supply the missing pieces. Near-duplicate scans treated as independent observations create spurious associations — the dermatology audit by Abhishek et al. found 13.47% of unique lesions duplicated across train/test partitions, silently inflating benchmark scores, and nothing stops the same failure mode in imaging cohorts. Detecting those duplicates perceptually — via short fingerprints that stay similar under reslicing, re-registration, and recompression — is exactly what perceptual hashing does, but at cohort scale the pairwise search cost explodes, which is where elastic partition scaling, in the style of Carousel, enters. The plan: a causal autoencoder for multimodal brain imaging whose training stream is first audited and deduplicated by perceptually hashed latents, with the hash search elastically partitioned to stay within memory, and the whole thing validated against longitudinal ground truth in the ADCB simulator.
Core Claim
The claim is specific and falsifiable. A multimodal autoencoder with a causal layer over its latent variables — oriented by an instrumental variable in the exogenous noise, following the Mendelian randomization logic that Shen et al. validated on UK Biobank — can correctly label individual latent deviations as upstream drivers or downstream consequences from cross-sectional data alone, but only if near-duplicate observations are first removed by a perceptual-hash audit; duplicates left in place will systematically bias the direction estimates, and the audit is only feasible at scale because hash buckets split and merge dynamically under occupancy feedback rather than being searched pairwise.
The obvious intersection of these four directions — making TOOD's latent-tree out-of-distribution detector scalable by dynamically partitioning its Hamming search, per the bottleneck its authors flag in Tree-based Ensemble Learning for Out-of-distribution Detection — is deliberately not the proposal. That fixes a compute problem while leaving the scientific question (what an out-of-distribution shift means) untouched. Here, elastic partition scaling is a supporting mechanism, perceptual hashing is a data-integrity instrument, the autoencoder is the representation core, and cross-sectional causal ambiguity is the actual subject.
Research Program
Three components get built. First, an imaging-augmented ADCB: the ADCB simulator already provides an expert-verified directed acyclic graph, a latent disease subtype, treatments, and 870 ADNI patients sampled at 12-month intervals — the only setting in the file where longitudinal ground truth exists for cross-sectional causal questions. We add a generative imaging head that renders synthetic multimodal volumes (an atrophy-like channel, an amyloid-like channel, a tau-like channel) from the latent disease variables, plus a nuisance process that emits near-duplicate views of the same subject: re-registered, recompressed, cropped variants that are not bitwise identical, mirroring the real leakage pattern that exact matching missed in DermaMNIST auditing. This gives controlled ground truth for both causal direction and duplicate contamination.
Second, the model: a normative multimodal autoencoder trained on simulated controls, in the spirit of Kumar et al., whose latent space carries a causal layer — an encoder producing exogenous noise $\epsilon$, transformed into endogenous latent factors by a structural equation $z = A^T z + \epsilon$ as in CausalVAE, but with the graph depth rules of VACA so that interventions propagate to all descendants. Unlike VACA, we do not assume causal sufficiency silently: a polygenic-risk-score-like instrument, available in the simulator by construction, regresses onto specific exogenous terms to orient edges, transferring the Mendelian randomization move from population-level tabular analysis to individual-level latent counterfactuals. A perceptual-loss term in training, following Improving Image Autoencoder Embeddings with Perceptual Loss, keeps the latents aligned with structural rather than pixel-level similarity, which matters because the audit and the causal layer must see the same geometry.
Third, the audit: each observation's latent code is binarized into a short perceptual hash $h$, and two observations are flagged as pseudoreplicates when their Hamming distance $d_H(h_i, h_j) \le \tau$, with one representative retained per cluster. The threshold $\tau$ is the one hand-tuned knob, calibrated on the simulator's known duplicate ground truth. The search is not pairwise. Hash prefixes define $2^k$ buckets, and $k$ is raised when bucket occupancy approaches the memory bound and lowered when buckets underflow — the feedback rule Carousel used to log 10,000 worm sources in 200 seconds where naive logging took 4,000, and the same principle SAME used to keep a video encryption queue from overflowing under bursts. Candidates are only compared within and between adjacent buckets, which keeps near-neighbor search subquadratic when duplicate bursts arrive, the regime where Carousel's adversarial testbed showed static schemes collapsing.
One design constraint deserves honesty: partitioning a causal model's variable set is dangerous. Dsep-CP showed that recursive variable partitioning creates "ghost edges" through d-separation violations, which is why we partition observations (hash buckets over data items), never variables. The causal graph stays whole; only the deduplication search is sharded.
Baselines
The do-nothing reference is cross-sectional association itself: compute latent deviations with the normative autoencoder of Kumar et al. and report which factors correlate with severity, without any causal layer or orientation. On Lacuna, no reported result was found for the Disease Severity Index's quantitative prediction strength, so this baseline is reimplemented on the extended ADCB and its associations reported on our scale; it anchors how much of the direction-recovery gap correlation alone can close (expected: the direction labels, none).
The strongest existing method is VACA, which is the only published causal autoencoder that answers interventional and counterfactual queries from cross-sectional observational data with a known graph. On the 11-node Adult semi-synthetic benchmark it reported an interventional maximum mean discrepancy of $12.68 \pm 1.95$, versus $155.52 \pm 5.93$ for MultiCVAE and $34.31 \pm 5.77$ for CAREFL — the number to beat on distribution-estimation metrics. We run VACA twice under an identical compute and tuning budget: once on the tabular ADCB covariates (its native modality) and once on our audited latent codes (its best shot at the imaging signal). Beating it means lower CATE error on the extended ADCB ground truth and higher edge-direction accuracy at matched sample size; matching its Adult-scale MMD on our harder imaging benchmark would itself be informative.
The longitudinal ceiling is ADCB's own sequential recurrent T-learner, which the benchmark showed reduces conditional average treatment effect error as history length grows while static single-time-point learners gain nothing from added context. At history length one it degenerates to a static T-learner, which is the fair comparison point for a cross-sectional method; with full history it is the upper bound we do not claim to reach. Two ablations isolate the proposal's ingredients: the full pipeline with the hash audit disabled (duplicates left in place), and the audit run with a fixed bucket count and brute-force pairwise fallback at equal compute — the first tests whether deduplication is load-bearing, the second whether the dynamic partitioning earns its complexity over a static hash table.
Why not just…?
Why not just collect longitudinal data and use the sequential learners ADCB already validates? This is the most serious objection, and the honest answer is that with full history the sequential learner will likely win — ADCB showed its error shrinking as history grows. But the entire premise of the cross-sectional ambiguity literature is that longitudinal imaging at UK Biobank scale takes years to accrue while cross-sectional snapshots exist now, in the tens of thousands. The proposal's decisive check is the head-to-head at history length one, where the sequential learner is by construction a static T-learner that ADCB showed cannot exploit context; if the instrumented causal autoencoder does not beat it there, the claim fails. One partial mitigation from the file: ADCB also found the sequential learner most vulnerable to collider bias when conditioning on current diagnosis, a failure mode latent-variable orientation avoids by never conditioning on post-treatment staging.
Why not just run Mendelian randomization on the tabular data, as Shen et al. already did? Where a valid genetic instrument exists, MR is a genuinely strong competitor for the direction question — it delivered the depression-to-white-matter orientation ($\beta$: 0.125 to 0.868, $p_{\text{FDR}} < 0.043$) that correlations could not. But MR yields one population-average arrow, not individual-level counterfactuals, not imaging-level mechanism, and nothing about which of a given patient's latent deviations are drivers. The proposal uses the instrument rather than competing with it: MR-on-tabular is included as a direction-recovery baseline, and the added value being tested is whether the same orientation logic, embedded in a latent causal layer, extends to per-subject multimodal counterfactual queries. If it recovers directions no better than MR while costing far more, that will be reported as a negative result.
Why not just deduplicate with exact or cryptographic hashes, or skip deduplication and brute-force the pairwise search? Exact matching provably misses the failure mode: the DermaMNIST duplicates that made up 13.47% of unique lesions differed in capture angle, framing, and compression, and were caught only by embedding-similarity auditing; cryptographic hashes avalanche under a single-pixel change (Robust Visual Fingerprinting for Large Scale Media Analysis), which is the same reason NFT asset verification needed perceptual rather than bitwise fingerprints. On brute force: the pairwise cost is the identical $O(M^2)$ bottleneck the TOOD authors flag for their Hamming evaluation, and Carousel's memory-bounded testbed is the direct evidence that occupancy-driven partition scaling handles bursty streams that overflow static schemes — 200 seconds versus 4,000 for 10,000 sources, and 300 versus 6,000 under adversarially synchronized traffic. The honest concession: at ADCB's 870 patients, pairwise search is cheap and dynamic partitioning contributes nothing; its value is only tested in the explicit scale experiment, and if static buckets match throughput at ten million items, that component is dropped from the claim.
Proposed Experiments
Experiment one measures direction recovery on extended ADCB. For every latent edge with known ground-truth orientation, each method — correlation-only, the normative autoencoder without a causal layer, tabular VACA, VACA on audited latents, and the full instrumented model — is scored on direction accuracy and CATE error across the simulator's tunable regimes (sample size, overlap, effect heterogeneity), using ADCB's own protocol so results are commensurable with its reported finding that sequential learners improve with history. Success is beating tabular VACA on direction accuracy while staying within the MMD scale it established on Adult ($12.68 \pm 1.95$).
Experiment two is the leakage ablation: sweep the injected duplicate rate from zero to 30% (bracketing DermaMNIST's 13.47%) and measure spurious-edge rate and CATE-error inflation with the audit on and off. The load-bearing prediction is that direction accuracy degrades measurably with duplicates when the audit is off and stays flat when it is on; if duplicates turn out not to bias the causal estimates, the audit is unnecessary and the core claim must be revised.
Experiment three is the scale test: time to complete duplicate-cluster coverage from $10^3$ to $10^7$ items, brute-force versus static buckets versus dynamic partitioning, including a synchronized adversarial burst schedule modeled on Carousel's worst case. The pass condition is subquadratic scaling with no missed clusters relative to the brute-force reference; failure reverts the system to the cheapest sufficient design.
Experiment four is the real-data probe: apply the full pipeline to actual ADNI multimodal imaging with a polygenic risk score as instrument, and ask whether any recovered orientation agrees with the MR-validated depression-to-white-matter direction where the domains overlap. This is framed as a consistency check, not proof — ADNI's depression phenotyping is thinner than UK Biobank's — but disagreement with an established instrumental result would be a serious red flag.
The falsification conditions are stated up front: if the instrumented latent model cannot beat a static T-learner on cross-sectional slices, or if duplicate contamination does not measurably distort causal estimates, the proposal's central claim is wrong and the experiments are designed to say so plainly.
References
- A phenome-wide association and Mendelian Randomisation study of polygenic risk for depression in UK Biobank
- Analyzing heterogeneity in Alzheimer disease using multimodal normative modeling on imaging‐based ATN biomarkers
- Variational Causal Autoencoder for Interventional and Counterfactual Queries
- CausalVAE: Structured Causal Disentanglement in Variational Autoencoder
- Investigating the Quality of DermaMNIST and Fitzpatrick17k Dermatological Image Datasets
- Carousel: Scalable Logging for Intrusion Prevention Systems
- Tree-based Ensemble Learning for Out-of-distribution Detection
- ADCB: An Alzheimer's disease simulator for benchmarking observational estimators of causal effects
- Improving Image Autoencoder Embeddings with Perceptual Loss
- A Speed-Adaptive Media Encryption Scheme for Real-Time Recording and Playback System
- Accelerating Recursive Partition-Based Causal Structure Learning
- Do NFTs’ Owners Really Possess their Assets? A First Look at the NFT-to-Asset Connection Fragility