Identity-Duplication Auditing in National-Scale Neuroimaging Repositories

1. INTRODUCTION

• The same person can appear under multiple identifiers when data are aggregated, a phenomenon termed identity duplication.Identifiers may be valid within individual datasets without being globally unique across them.Source: §1 · p. 2
• Privacy, version-control, and release-management practices can interact at repository scale to create identity duplication.The paper characterizes this as a structural risk emerging from otherwise valid data-governance processes.Source: §1 · p. 2
• Identity duplication can create train-test leakage, inflate apparent model performance, and increase false-positive risk in statistical analyses.Repeated scans from one person can appear in both training and test sets, weakening evaluation of generalization to new participants.Source: §1 · p. 3
• HAPPEN combines SHA-256 detection of identical scans with learned retrieval of non-identical candidate scans, which undergo human review.The paper presents this as an end-to-end workflow for auditing identity duplication.Source: §1 · p. 3
• The workflow was evaluated across all 95,129 scans in the national-scale repository and deployed externally without model retraining.The external deployment was at the University of Pennsylvania.Source: §1 · p. 4
• HAPPEN extends MRI subject matching from pairwise retrieval to a reviewed audit that reconciles records into subject-level groups.The paper describes evaluation using genetic evidence and external deployment without transferring imaging data.Source: §1 · p. 4

2. RELATED WORK

• NIMH’s Data Archive developed a repository-wide GUID system; NACC uses NACCID to link all available data; UK Biobank assigns encrypted identifiers unique within an access application; these mechanisms are typically project-scoped [10],,.Source: §2 · p. 4
• MRI-based approaches have used structural morphology, learned representations, and 3D keypoint signatures for subject matching and curation,,.The paper states that these studies establish MRI's utility for subject matching.Source: §2 · p. 4
• Prior MRI methods are described as stopping at retrieval or pairwise flags rather than providing a repository-wide identity audit.HAPPEN treats pairwise matching as an intermediate result and sends candidate links for human review before reconciling groups.Source: §2 · p. 5
• HAPPEN's repository-scale evaluation covered 95,129 scans from 21 datasets and assessed sensitivity using 54 genetically confirmed duplicate pairs.An independent institution also ran the workflow without retraining or transferring imaging data.Source: §2 · p. 5

3.1 Stage I: exact-duplicate detection

• Stage I defines exact duplicates as identical voxel arrays after noninterpolating RAS+ reorientation and groups scans using SHA-256 fingerprints computed from canonical image bytes, not headers.This definition covers identical voxel content despite differences in compression, header metadata, or stored orientation; intensity-altered scans fall outside Stage I.Source: §3.1 · p. 6

3.2 Stage II: near-duplicate detection and human-in-the-loop review

• Stage II embeds scans with supervised contrastive learning, retrieves 500 cosine-similar neighbors via FAISS, filters same-record matches, and sends thresholded cross-record pairs for Yes/No/Maybe review.Embedding similarity identifies candidates but does not establish that scans came from the same individual.Source: §3.2 · pp. 6–7

3.3 Duplicate reporting

• Reviewed Yes pairs connect repository subject records into near-duplicate groups, and records linked by Stage I or Stage II are consolidated into final groups.No and Maybe decisions do not contribute to subject-level grouping.Source: §3.3 · p. 7
• HAPPEN reports exact duplicates across four provenance levels and near duplicates within or across datasets, using interactive matrices to summarize repository-wide distributions.The provenance levels describe relationships within a group and are not mutually exclusive group labels.Source: §3.3 · pp. 7–8

3.4 Implementation and reproducibility

• HAPPEN reuses completed hashes and embeddings for unchanged scans, resumes interrupted audits, and supports portable execution through Docker or Apptainer containers.Only new or modified scans require preprocessing and embedding inference.Source: §3.4 · p. 8
• Audit configuration is saved with each run, stochastic components use fixed random seeds, and source code and documentation are available.The interface is configured through a TOML file.Source: §3.4 · p. 8

3.5.1 Data sources

• The study repository comprised 21 datasets with 40,218 recorded subjects and 95,129 T1-weighted brain MRI scans; BLSA and VMAP were withheld, while the remaining 19 datasets supported development and internal evaluation.An independent UPenn repository contained 6 datasets, 14,690 recorded subjects, and 22,386 T1-weighted brain MRI scans.Source: §3.5.1 · p. 8

3.5.2 Duplicate-aware construction of the development and test datasets

• To limit inflated performance, the team removed exact duplicates and unusable scans, screened near-duplicate candidates with PCA and visual review, and split development data by subject.The split was 8:1:1 for training, validation, and internal testing, with every scan from a subject kept in the same split.Source: §3.5.2 · pp. 9–10
• The development subset included 5,000 subjects with at least three non-identical scans and 457 subjects with exactly two scans.Training views were balanced at six per subject; validation and test scans were not augmented.Source: §3.5.2 · pp. 9–10

3.5.3 Evaluation of Stage I and Stage II

• Stage I controls tested grouping of byte-identical, metadata-modified, and orientation-modified copies while separating scans altered by one voxel.The protocol used 100 randomly selected scans and expected the three positive controls to group with their source scan.Source: §3.5.3 · p. 11
• Stage II retrieval was evaluated on the internal test split, BLSA, and VMAP using all scans as queries and APR, mAP, and ROC-AUC.Same-record scans were positives and different-record scans negatives; operating-threshold selection set the similarity cutoff for retaining retrieved pairs as human-review candidates.Source: §3.5.3 · p. 11
• Stage I and Stage II implementation evaluation used two input configurations, including 3D-SIFT-FS and 3D-SIFT-Shared, and assessed computational efficiency on 200 scans using 32 workers.Efficiency measures were median representation time with IQR, mean search time per query, and median representation size; training and preprocessing time were excluded.Source: §3.5.3 · p. 11
• When large-scale cross-dataset identity labels were unavailable, longitudinal pairs from the same repository subject record served as known same-subject references for threshold selection.The design framed lower thresholds as retaining more potential matches but increasing review workload, while higher thresholds reduce review workload at the risk of missed matches.Source: §3.5.3 · p. 12
• Across 12 thresholds from 0.85 to 0.96, the study compared same-subject-pair retention and a review-efficiency proxy to select a practical default for repository deployment.The review-efficiency proxy measured the concentration of same-subject pairs in the simulated review set.Source: §3.5.3 · p. 12

3.5.4 End-to-end pipeline evaluation

• End-to-end deployment assessment covered wall time, cumulative CPU time, GPU inference time, human-review time, stage-specific findings, and whether duplication was within or across datasets.Stage I findings covered exact-duplicate groups and scans; Stage II findings covered candidate pairs, review decisions, and subject-level near-duplicate groups.Source: §3.5.4 · p. 12
• 54 genotype-defined cross-dataset pairs (PI_HAT > 0.9) were used to measure sensitivity as the proportion placed together in final subject groups, with a 95% Wilson confidence interval.None of the reference subjects was included in the Stage II training split.Source: §3.5.4 · p. 12
• The external deployment used a frozen encoder without retraining; UPenn independently selected datasets, assigned a local reviewer, and used review criteria that were not harmonized with the study deployment.The external reviewer applied a more conservative standard, so the deployments were not designed as a comparison of model precision.Source: §3.5.4 · pp. 12–13
• UPenn set the operating threshold to 0.99 according to its local review budget, and all MRI data remained within its environment throughout deployment.The job used 32 CPU cores, an NVIDIA RTX A6000 GPU, and 32 parallel processes.Source: §3.5.4 · p. 13

4.1 Stage I evaluation

• All 300 positive controls shared their source scan’s fingerprint, while all 100 one-voxel-modified controls received a different fingerprint.Source: §4.1 · p. 13
• Stage I was invariant to metadata and stored-orientation changes when canonical voxel content was preserved, but sensitive to changes in voxel values.Source: §4.1 · p. 13

4.2 Stage II evaluation

• HAPPEN’s reported rows list APR values of 0.9975, 1.0000, and 0.9956; mAP values of 0.9997, 1.0000, and 0.9975; and ROC-AUC values of 1.0000, 1.0000, and 0.9999.The table identifies APR, mAP, and ROC-AUC as its retrieval metrics.Source: §4.2 · p. 13, Table 2 · p. 14

Panel B. Computational efficiency

• HAPPEN outperformed the PCA baseline on all three metrics; it achieved the highest APR in the internal test and BLSA cohorts and matched VMAP ceiling performance.Shared preprocessing improved SIFT performance and narrowed its difference from HAPPEN, particularly for mAP and ROC-AUC.Source: Panel B. Computational efficiency · p. 14
• At λ = 0.92, the selected threshold retained more than 98% of known same-subject pairs while avoiding the sharper review-efficiency loss seen at lower thresholds.It was chosen as a practical operating point, not a universally optimal threshold.Source: Panel B. Computational efficiency · p. 14

4.3.1 Deployment in the study repository

• In the study repository, Stage I identified 1,316 exact-duplicate groups involving 2.8% of all scans, while 1,275 subject-level groups involved 6.4% of recorded identifiers; 82.4% linked records across datasets.Source: §4.3.1 · p. 15
• 100% sensitivity recovered all 54 available genetic-reference pairs in the evaluation.The 95% confidence interval was 93.4%-100%.Source: §4.3.1 · p. 15

4.3.2 External-site repository deployment

• Fewer than 0.1% of external-site scans involved exact duplication, while reviewers judged 9.1% of Stage II candidate pairs likely near duplicates; 97.6% of resulting groups spanned datasets.The external review standard was more conservative, so the candidate-pair fraction is not directly comparable with the study repository.Source: §4.3.2 · p. 16

4.3.3 Deployment performance

• The study-repository deployment finished in under seven days, and GPU inference time remained small relative to cumulative CPU time.The two sites used different computing environments, so their runtimes are not directly comparable.Source: §4.3.3 · p. 16

5. DISCUSSION

• Stage II exceeded PCA and 3D-SIFT-FS in the internal test and BLSA cohorts, with comparable or higher retrieval point estimates than 3D-SIFT-Shared.Performance was near-ceiling in the internal test and two withheld cohorts.Source: §5 · pp. 16–17
• HAPPEN’s fixed-length embedding was faster to generate and search and smaller than the variable-length SIFT fingerprint; fewer than 10,000 candidate pairs required 4.5 hours of review.The authors suggest reducing preprocessing steps and distributing processing across multiple machines in future versions.Source: §5 · pp. 17–18
• The complete workflow recovered duplication cases identified by genome-wide genotype data.The discussion presents this genetic-reference evaluation alongside the retrieval results.Source: §5 · p. 17
• The authors frame identity duplication as a repository quality-assurance concern and describe uses in repository releases and participant-level analysis.Grouping linked scans can help keep each individual within a single training, validation, or test set.Source: §5 · p. 17
• Exact-duplicate groups involved 2.8% of all scans, near-duplicate groups involved 6.4% of recorded identifiers, and 82% of near-duplicate groups crossed datasets.Duplicate groups were found both within datasets and across dataset boundaries.Source: §5 · p. 17
• UPenn independently completed the full workflow without imaging-data transfer, choosing its own higher similarity threshold and review policy.This demonstrated end-to-end operation outside the development environment; local choices can balance sensitivity, candidate volume, and review effort.Source: §5 · p. 17
• Interpretation is limited by possible monozygotic-twin ambiguity and Stage II’s evaluation only on T1-weighted MRI.The authors suggest pedigree or source identity records for ambiguous cases and evaluation on other modalities and challenging scans.Source: §5 · p. 18
• The study defines identity duplication as one person recorded under multiple identifiers; the reverse case—one identifier containing multiple people—remains a future task.A future version could compare scans within each identifier and flag inconsistent groups for review.Source: §5 · p. 18

• Jiheng Li is credited with conceptualization, methodology, software, validation, formal analysis, investigation, data curation, supervision, project administration, funding acquisition, visualization, and original-draft writing.Source: CRediT AUTHORSHIP CONTRIBUTION STATEMENT · p. 18
• Michael E. Kim is credited with conceptualization, data curation, manuscript review and editing, and supervision.Source: CRediT AUTHORSHIP CONTRIBUTION STATEMENT · p. 18
• Trent M. Schwartz and Gaurav Rudravaram are credited with supervision; Yuhan Cui with investigation, resources, and data curation; and Derek B. Archer with resources, data curation, and funding acquisition.Source: CRediT AUTHORSHIP CONTRIBUTION STATEMENT · p. 18
• Timothy J. Hohman is credited with resources, data curation, and funding acquisition; Lori L. Beason-Held and Dario J. Englot with resources; and Victoria L. Morgan with resources, review and editing, and funding acquisition.Source: CRediT AUTHORSHIP CONTRIBUTION STATEMENT · p. 18
• Angela L. Jefferson is credited with resources and funding acquisition; Lianrui Zuo with supervision and review and editing; Guray Erus and Christos Davatzikos with supervision and project administration; and Bennett A. Landman with conceptualization and review and editing.Source: CRediT AUTHORSHIP CONTRIBUTION STATEMENT · p. 18

• Timothy J. Hohman serves on the scientific advisory board of Circular Genomics and as an editor for two Alzheimer’s & Dementia journals.Source: DECLARATION OF COMPETING INTEREST · p. 18
• The remaining authors declare no known competing financial interests or personal relationships that could have influenced the reported work.Source: DECLARATION OF COMPETING INTEREST · p. 18

• The study’s participant-level data were obtained under applicable repository access procedures and data-use agreements.Source: DATA AVAILABILITY · p. 18
• The authors are not authorized to redistribute the aggregated study repository.Source: DATA AVAILABILITY · p. 18
• External-site participant-level data remain under University of Pennsylvania control and are not available from the authors.Source: DATA AVAILABILITY · p. 18

• The authors used OpenAI ChatGPT to assist with language editing and stylistic refinement during manuscript preparation.Source: DECLARATION OF GENERATIVE AI AND… · p. 23
• The authors reviewed and revised all suggestions and take full responsibility for the work.Source: DECLARATION OF GENERATIVE AI AND… · p. 23

S1. Repository composition

• BLSA and VMAP were withheld evaluation cohorts; the other 19 datasets supported model development and internal evaluation, with WRAP and WRAPnew combined.The repository comprised 21 datasets, and eligibility for the longitudinal-rich model-development pool required at least three eligible T1-weighted scans per subject.Source: S1. Repository composition · pp. 27–28

S2. Stage II model development

• Stage II treated repeated scans from one repository subject record as positive views and scans from different records as negatives.The main-manuscript preprocessing workflow was used during training and inference; PCA screening and augmentation were restricted to model development.Source: S2. Stage II model development · p. 28

S2.1 PCA near-duplicate screening

• Incremental PCA screening pooled development, BLSA, and VMAP scans before partitioning; a researcher reviewed high-cosine-similarity pairs across subject records and removed judged near duplicates.The cleaning screen used 352-scan batches and 176 components, and was distinct from the later PCA retrieval baseline fitted on the training split.Source: S2.1 PCA near-duplicate screening · pp. 28–29

S2.2 Training-view construction and augmentation

• Training augmentation produced six positive views per subject, balancing three-scan subjects with three original and three augmented views and two-scan subjects with two original and four augmented views.Augmentation was applied only to the training split.Source: S2.2 Training-view construction and… · p. 29

S2.3 Input representation, architecture, and objective

• Inputs used the native 1-mm MNI152NLin2009cSym grid, were center-cropped or zero-padded to 160 × 192 × 160 voxels, and retained 32 central axial slices for the encoder.The encoder used a 3D ResNet-like architecture with squeeze-and-excitation channel attention.Source: S2.3 Input representation,… · p. 29
• The L2-normalized 512-dimensional representation supported both supervised contrastive training and near-duplicate retrieval; the loss compares an anchor with same-subject views and uses τ=0.1.The representation followed global average pooling and a two-layer projection MLP with LayerNorm.Source: S2.3 Input representation,… · pp. 29–30

S2.4 Training hyperparameters

• Training used 100 epochs without early stopping, AdamW, and a OneCycle learning-rate schedule with warm-up over 10 epochs and cosine decay over the remaining 90.The schedule decayed to 1 × 10^-8; mixed precision with GradScaler was also used.Source: S2.4 Training hyperparameters · p. 30

S3. Baseline and efficiency-benchmark implementation

• Efficiency was measured on 200 scans with 32 workers after a 10-scan warm-up, reporting representation time, per-query search time, and representation size; preprocessing and training time were excluded.HAPPEN and PCA used flat FAISS search, while 3D SIFT-Rank used its official all-to-all matcher.Source: S3. Baseline and efficiency-benchmark… · p. 30