PaperlayerView exact revision on arXiv

Source-linked AI summary

NEXUS-MI: Communication-Aware Federated Personalization for Gateway-Coordinated Motor-Imagery Brain-Computer Interfaces

Daniel Adu Worae, Aarthy Nagarajan

arXiv:2609.09786v1cs.LGcs.NI

TL;DR

MI-BCI personalization must handle subject and session variation while federated learning must operate under irregular gateway connectivity. NEXUS-MI evaluates synchronization as a learning-and-communication control problem and finds substantial traffic savings but small, realization-dependent cohort accuracy differences and important subject-level vulnerability.

  • Problem

    Existing federated MI-BCI evaluations largely assume regular synchronization, leaving communication policy, personalized decoding, and subject-level reliability under heterogeneous connectivity insufficiently characterized.

  • Method

    NEXUS-MI keeps raw EEG and classifier heads at gateways while an edge coordinator manages a shared backbone across six heterogeneous-link policies, using session-based offline replay and a controlled P3/P5 comparison.

  • Results

    Communication-aware coordination reduced server-to-client backbone traffic by approximately 42% across five matched realizations, while corresponding accuracy differences were small, realization-dependent, and statistically uncertain.

  • Takeaways & Limitations

    Federated MI-BCI evaluation should jointly report personalized accuracy, synchronization cost, update freshness, and subject-level reliability because cohort averages can conceal user-specific losses.

  • Takeaways & Limitations

    The findings are limited to offline replay on two public session-based datasets from healthy participants under a one-subject/one-gateway abstraction.

Abstract

from arXiv · show

Electroencephalography (EEG)-based motor-imagery brain-computer interfaces (MI-BCIs) vary across subjects and sessions, complicating personalization from limited calibration data. Federated learning can exploit shared representations without centralizing raw EEG, but existing federated MI studies largely assume regular synchronization. We introduce NEXUS-MI, a gateway-coordinated federated personalization framework that treats synchronization as a coupled learning-and-communication control problem. Raw EEG and classifier heads remain local, while an edge coordinator maintains the shared backbone. We evaluate NEXUS-MI through offline replay using BCI Competition IV Dataset 2a (BCICIV-2a; 9 subjects, 4 classes) and OpenBMI (54 subjects, 2 classes). Session 1 supports backbone learning, and Session 2 provides limited-calibration personalization and held-out testing. An ideal-link reference and six heterogeneous-link policies characterize gateway participation, buffering, stale-update admission, and backbone-download control. The principal comparison holds delayed-update handling fixed while contrasting non-adaptive and communication-aware synchronization. Paired subject-level comparisons use Holm adjustment, and robustness across five matched realizations is assessed by hierarchical bootstrap. Communication-aware coordination reduced server-to-client backbone traffic by approximately 42% on both datasets, while cohort-level accuracy differences were small and realization-dependent. Cohort averages also concealed subject-level vulnerability, with losses reaching approximately 12 percentage points on BCICIV-2a relative to the ideal-link reference. These findings establish gateway synchronization as an explicit design variable in federated MI personalization and motivate joint evaluation of personalized accuracy, communication cost, update freshness, and subject-level reliability.

I. INTRODUCTION

NEXUS-MI addresses unstable, user-dependent MI-BCI personalization by treating gateway synchronization as part of the learning system. It evaluates communication policies across heterogeneous links, limited-calibration sessions, and subject-level reliability rather than mean accuracy alone.

  • Motivation: MI-BCI decoders vary across subjects, sessions, and recording conditions, making reliable personalization from limited calibration difficult.Relevant factors include physiology, electrode placement, attention, fatigue, and session drift.
  • Motivation: Existing approaches trade off small-set adaptation, population-level learning, and privacy without fully addressing deployment-time synchronization.Subject-specific training may overfit, while centralized training moves sensitive EEG away from the user environment.
  • Framework: NEXUS-MI treats gateway participation, buffering, stale-update admission, and backbone-download control as determinants of personalized decoding.The framework keeps raw EEG and subject-specific heads local while the edge coordinator maintains the shared backbone.
  • Evaluation: The evaluation replays two public session-based datasets, using earlier-session data for shared learning and later-session data for limited-calibration personalization and held-out testing.BCICIV-2a contains nine subjects and four classes; OpenBMI contains 54 subjects and two classes.
  • Results: Across five matched realizations, communication-aware coordination reduced server-to-client backbone traffic by approximately 42% on both datasets, while accuracy differences were small and realization-dependent.The P5–P3 accuracy differences had hierarchical-bootstrap 95% confidence intervals that included zero.
  • Reliability: Subject-level analysis measures worst-subject performance, threshold violations, deviations from the ideal-link reference, and sensitivity to link availability.These measures test whether aggregate accuracy conceals uneven effects across users.

B. Federated and Personalized Learning for EEG and BCI

The paper positions NEXUS-MI at the intersection of federated EEG learning, personalized shared-backbone models, and communication-aware edge coordination. Its deployment framing makes synchronization behavior an explicit factor in MI-BCI personalization while retaining local data and heads.

  • Federated EEG and BCI Learning: Federated EEG learning exploits cross-subject information without centralizing raw recordings, but much prior work frames federation primarily as distributed model training.This leaves gateway participation and synchronization behavior outside much of the evaluation scope.
  • Personalized Learning: Personalized federated learning separates shared representations from local heads, aligning with MI-BCI needs for reusable EEG structure and limited-calibration adaptation.NEXUS-MI adopts this shared-backbone and subject-specific-head structure.
  • Communication-Aware FL: Communication-efficient federated systems address participation, heterogeneity, local computation, aggregation, and communication cost, but commonly emphasize convergence or aggregate performance.These objectives do not by themselves establish reliable feedback for users with weak links.
  • Communication-Aware FL: Gateway coordination can alter which updates arrive, when stale contributions are admitted, and which backbone version supports subsequent personalization.Communication mechanisms therefore function as determinants of MI-BCI personalization quality, not only as systems optimizations.
  • NEXUS-MI Positioning: NEXUS-MI evaluates synchronization using session-based personalization accuracy, update cost, backbone synchronization, and related deployment outcomes.The framework applies these mechanisms while raw EEG, labels, and personalized heads remain at gateways during personalization and synchronization.
  • Scope: The evaluation is a replay-based deployment model over public session-based datasets rather than a real-time wearable implementation.Formal privacy mechanisms are not evaluated, and pooled Session-1 data initialize EIB-PH before gateway-coordinated personalization.

B. Session-Based Personalization Protocol

The protocol separates cohort-level backbone learning from subject-local later-session adaptation, while gateway communication determines the backbone state and update freshness available during synchronization.

  • Session-Based Personalization Protocol: Earlier-session data support shared representation learning and an initial subject-specific decoder state; later-session data provide limited calibration and held-out testing.This protocol evaluates continued-use personalization rather than only training-time classification.
  • Session-Based Personalization Protocol: Each subject’s later-session data are divided into a limited calibration set and a disjoint held-out evaluation set.Held-out trials are used only to measure later-session decoding after personalization.
  • Backbone–Head Decomposition: Each local decoder combines a shared EEG feature backbone with a subject-specific classifier head.The backbone captures representations benefiting from multi-subject learning, while the head represents the subject-specific decision boundary.
  • Backbone–Head Decomposition: During Session 1, gateways jointly optimize the backbone and head, but during Session 2 they hold the collaborative backbone fixed and adapt only the retained subject-specific head.Session-2 personalization starts from a head learned during Session 1 rather than a random initialization.
  • Gateway-to-Edge Communication Model: Gateway availability determines whether local operation can upload updates or download refreshed backbones in a synchronization round.Unavailable gateways may continue using cached model states while communication is absent.
  • Gateway-to-Edge Communication Model: Intermittent communication creates version lag and update-freshness differences between coordinator and gateway backbones.The coordination policy acts on availability, versions, pending updates, and participation to govern synchronization.

E. Problem Statement

NEXUS-MI formulates federated MI-BCI personalization as a gateway-coordination problem that jointly shapes personalized accuracy, communication cost, update freshness, and subject-level reliability. It separates shared backbone learning from local classifier-head adaptation and evaluates these mechanisms under limited later-session calibration.

  • Each policy induces a subject-level accuracy and communication operating point under fixed data, learning regime, and calibration budget.The formulation tracks held-out later-session accuracy, client-to-server traffic, server-to-client traffic, rejection, accepted-update lag, and buffered-upload delay.
  • NEXUS-MI treats gateway coordination as part of the federated learning system rather than as independent communication overhead.Policies are compared as alternative operating points because communication savings, rejection, staleness, and delay can affect subjects differently.
  • The backbone captures cohort-level EEG representations, while the classifier head represents the subject-specific decision boundary.This decomposition supports shared representation learning with local subject adaptation.
  • SB-PH: SB-PH learns the backbone collaboratively and personalizes only each gateway’s classifier head after collaborative training.The regime tests whether a cohort-learned representation supports limited-calibration adaptation while retaining a local decision boundary.
  • EIB-PH: EIB-PH initializes the backbone with pooled Session 1 representation learning before gateway-coordinated training and later-session head personalization.The subsequent gateway-local structure matches SB-PH, with raw EEG, labels, and subject-specific heads retained locally.
  • Reporting SB-PH and EIB-PH separately distinguishes representation initialization from gateway-coordination effects.For both regimes, later-session personalization starts from the final collaborative backbone and updates only the subject-specific head using limited calibration data.

B. Synchronization Conditions

NEXUS-MI compares an unconstrained ideal-link reference with heterogeneous-link policies that vary gateway participation, buffering, stale-update admission, scheduling, and backbone downloading. The principal controlled comparison holds delayed-update handling fixed while contrasting non-adaptive and communication-aware synchronization.

  • The ideal-link reference removes communication-induced constraints while retaining the same collaborative personalization pipeline.Gateways synchronize when required, upload current-round updates, and avoid missed uploads, delayed delivery, stale rejection, and checkpoint unavailability.
  • Heterogeneous connectivity assigns gateways to high-, moderate-, or low-availability groups with independently sampled online states across rounds.Matched policy comparisons reuse the same availability trace, while offline gateways retain cached state and may buffer updates when enabled.
  • The default availability setting uses online probabilities 0.95, 0.70, and 0.40 for high-, moderate-, and low-availability gateways.A sensitivity study repeats the grouping under mild probabilities 0.98/0.85/0.60 and severe probabilities 0.90/0.50/0.20.
  • The six heterogeneous-link policies differ across scheduling, offline-update handling, stale-update admission, and backbone-download control.These dimensions define how gateways are selected, how missed updates are retained, which delayed updates are admitted, and when newer backbones are downloaded.
  • P1–P4 form a non-adaptive reference family, whereas P5–P6 adapt scheduling to currently online gateways and use stale-aware downloading.Within the non-adaptive family, differences isolate missed updates, buffering, and stale admission; P5 and P6 differ in offline-update retention.
  • P3 and P5 provide the principal controlled comparison because both use FIFO buffering and stale-update rejection.P3 uses non-adaptive all-gateway scheduling with always-download synchronization, while P5 uses communication-aware scheduling with stale-aware downloading.

E. Communication-Aware Gateway Scheduling

Communication-aware scheduling observes gateway availability before assigning immediate upload opportunities. It selects from online gateways using a fixed priority heuristic that favors synchronization need, buffered work, and local-backbone lag.

  • P5 and P6 determine online and offline gateway sets before selecting upload participants in each communication round.Only online gateways are eligible for immediate upload under the online-first rule.
  • The selection budget is K = 6 for BCICIV-2a and K = 40 for OpenBMI.If the online set does not exceed K, all online gateways are selected; otherwise, the highest-priority online gateways are chosen.
  • The fixed priority ordering favors gateways that have not uploaded recently, carry more urgent buffered updates, or hold locally stale backbones.Seeded tie-breaking and subject identifiers preserve reproducibility in matched comparisons.
  • The priority vector combines coordinator-version lag, time since successful upload, pending-buffer status, buffered-update staleness and delay, and local-backbone lag.The coordinator sorts online gateways in descending lexicographic order, with earlier components dominating later ones.
  • Under non-adaptive policies, every gateway is scheduled before availability is applied, whereas communication-aware policies assign immediate opportunities only to selected online gateways.Scheduled offline gateways train from cached backbones, with their updates discarded or retained according to the active offline-update rule.

F. Buffering, Checkpoints, and Stale-Update Admission

NEXUS-MI manages delayed updates through buffering, versioned checkpoint reconstruction, and stale-update admission, while download control limits unnecessary backbone transfers. Accepted fresh and buffered payloads are reconstructed and aggregated uniformly at the update-instance level.

  • Buffering: Offline-update rules either discard missed updates, retain them in bounded FIFO buffers, or replace pending updates with the latest update.FIFO buffering drops the oldest pending update when capacity is exceeded.
  • Checkpoints: Each communicated backbone delta carries metadata identifying the coordinator version from which local training began.The coordinator retains versioned checkpoints so delayed deltas can be reconstructed relative to their training base.
  • Checkpoints: Checkpoint retention and stale-update admission are separate controls: retention enables reconstruction, while the stale threshold determines admissibility.The checkpoint-retention margin is fixed at ρ = 5 across experiments and policies.
  • Stale-update admission: Delayed uploads are first rejected when their base checkpoint is unavailable; otherwise, their staleness is computed from coordinator and training versions.Stale-drop admission further rejects updates exceeding the allowed version lag.
  • Stale-update admission: The stale-update threshold is τs = 2, allowing modest delayed delivery while rejecting updates trained from excessively old backbone versions.Accepted delayed updates are reconstructed from their retained base checkpoint and transmitted backbone delta.
  • Aggregation: Accepted fresh and buffered payloads are independently reconstructed and uniformly averaged at the update-instance level.A gateway may contribute multiple accepted candidate backbone states in one round, while an empty accepted set leaves the current backbone unchanged.
  • Download control: Always-download refreshes a gateway whenever the coordinator has a newer backbone, whereas stale-aware downloading refreshes only after local lag exceeds τd = 1.This rule reduces unnecessary server-to-gateway transfers while preventing training at the stale-update boundary.

V. EXPERIMENTAL EVALUATION

The evaluation tests gateway-coordinated federated MI-BCI personalization under ideal and heterogeneous synchronization, using session-based limited-calibration protocols on two public datasets. It measures accuracy, communication, synchronization behavior, and subject-level reliability against an ideal-link reference.

  • Experimental protocol: The evaluation compares an ideal-link reference with six heterogeneous-link policies that vary participation, buffering, stale-update handling, and backbone-download control.The ideal-link condition provides a synchronization reference, while heterogeneous policies define distinct accuracy–communication operating points.
  • Experimental protocol: The experiments use BCICIV-2a and OpenBMI with Session 1 for shared-backbone learning and Session 2 for limited-calibration personalization and held-out testing.BCICIV-2a contains 9 subjects and four classes; OpenBMI contains 54 subjects and two classes.
  • Metrics: Accuracy is the primary learning outcome, complemented by communication traffic, rejected uploads, update staleness, buffered-upload delay, avoided downloads, and subject-level reliability.Reliability includes absolute accuracy, deviations from the ideal-link reference, low-accuracy counts, and large per-subject changes.
  • Statistical methodology: Statistical comparisons average calibration-budget results within subject, use paired subject-level differences, and report confidence intervals, effect sizes, and paired tests.Holm adjustment is applied across the four primary P5-versus-P3 paired t-tests.
  • Statistical methodology: Robustness is assessed across five matched training and gateway-availability realizations with a crossed hierarchical bootstrap preserving within-subject policy pairing.Communication outcomes use each matched realization as the replicate unit.
  • Ideal-link reference: Under ideal synchronization, OpenBMI EIB-PH reaches 80.75% mean accuracy versus 72.23% for SB-PH, while BCICIV-2a reaches 67.77% for SB-PH and 66.82% for EIB-PH.The ideal-link condition is a reference synchronization trajectory rather than an accuracy ceiling.

D. Accuracy and Communication under P1–P6

The P1–P6 evaluation shows that synchronization policies create coupled accuracy–communication operating points rather than bandwidth changes independent of learning. P5 consistently reduces server-to-client backbone traffic, while accuracy effects depend on dataset and personalization regime.

  • Cross-policy results: P5 and P6 reduce OpenBMI server-to-client backbone traffic to 21.98 MB, versus 37.49 MB under P3 and 54.56 MB under ideal synchronization.Within OpenBMI EIB-PH, P5 improves mean accuracy from 80.13% under P3 to 81.12%.
  • Cross-policy results: On BCICIV-2a, P4 achieves the highest mean accuracy, while P5 and P6 reduce server-to-client traffic to 3.68 MB with regime-dependent accuracy.P5 reaches 68.44% under EIB-PH and 66.77% under SB-PH; P6 reaches 68.44% and 64.51%, respectively.
  • Interpretation: Figure 2 treats each policy as a coupled learning-and-synchronization outcome because scheduling, buffering, stale-update admission, and download decisions alter update sequence and recency.Thus, lower downlink traffic does not necessarily imply an accuracy-neutral operating point.
  • P3/P5 comparison: P5 reduces server-to-client backbone traffic by 41.38% on OpenBMI and 41.34% on BCICIV-2a relative to P3.The corresponding savings are 15.51 MB on OpenBMI and 2.59 MB on BCICIV-2a.
  • P3/P5 comparison: On OpenBMI, P5 remains above P3 across calibration budgets, whereas on BCICIV-2a their accuracy relationship varies by calibration budget and learning regime.The communication reduction is consistent, but the accuracy effect depends on dataset and learning regime.
  • P3/P5 comparison: The strongest accuracy result is OpenBMI EIB-PH, where P5 increases mean accuracy from 80.13% to 81.12%, a +0.99 percentage-point difference with Holm-adjusted p = 0.013.This is the only Holm-adjusted accuracy difference reported as statistically significant in the primary comparison.

F. Component Analysis of P5

The component analysis attributes P5’s communication savings primarily to stale-aware backbone downloading and its accuracy advantage to the combination of download control and gateway scheduling. The full combination preserves the downlink reduction while achieving the strongest observed accuracy among tested variants.

  • Communication component: Stale-aware backbone downloading reduces OpenBMI EIB-PH server-to-client traffic from 37.49 MB to 21.98 MB, matching P5’s 41.38% reduction.Online-random and online-priority scheduling alone retain traffic near the P3 reference, with reductions of only 0.56%.
  • Accuracy component: Online-priority scheduling raises mean OpenBMI EIB-PH accuracy to 80.67%, compared with 80.37% for online-random scheduling and 80.43% for stale-aware downloading alone.The results indicate distinct contributions from gateway selection and backbone-download control.
  • Robustness: The robustness analysis repeats the P3/P5 comparison across five matched training and gateway-availability realizations under the default heterogeneous-link profile.The repeated analysis evaluates whether the principal operating point depends on one realization.

G. Sensitivity to Heterogeneous Link Availability

Communication-aware coordination consistently reduces server-to-client backbone traffic as gateway availability worsens, but accuracy effects vary by dataset and availability profile. Subject-level outcomes show why communication savings must be evaluated alongside reliability and shared-representation effects.

  • Synchronization effects: As availability becomes more constrained, coordinator-rejected uploads and buffered-upload delays increase, producing a more delayed and selective update stream.Accepted-update staleness remains policy- and profile-dependent, with higher staleness under P5 in the reported comparison.
  • Communication effects: P5 reduces server-to-client backbone traffic across mild, default, and severe availability on both datasets.On OpenBMI, traffic falls from 44.36 to 24.39 MB under mild availability and from 28.99 to 17.81 MB under severe availability; BCICIV-2a shows the same pattern.
  • Accuracy effects: P5 accuracy varies across profiles: it is nearly tied with P3 under mild OpenBMI availability, higher under default availability, and lower under severe availability.The default OpenBMI comparison is 81.12% versus 80.13% with Holm-adjusted p = 0.013; mild and severe confidence intervals include zero.
  • Subject-level reliability: Cohort-level gains can conceal subject-level vulnerability, including minimum differences of -12.04 percentage points for P5 on BCICIV-2a and -7.22 percentage points for P5 on OpenBMI.On OpenBMI, P5 also leaves 12 subjects below the 70% analysis threshold; on BCICIV-2a, P5 has a 21.30-percentage-point subject-level difference range.
  • Interpretation: Gateway coordination changes the shared-representation trajectory because it determines which updates arrive, when delayed updates are admitted, and which backbone version supports local personalization.Thus, communication policy is coupled to learning rather than operating only as a post-training bandwidth-control layer.

VII. CONCLUSION

NEXUS-MI frames gateway synchronization as part of federated MI-BCI learning rather than as a separate communication layer. The study finds a persistent traffic benefit but no uniformly dominant policy, so deployment evaluation must combine communication efficiency with personalized reliability.

  • Gateway synchronization helps determine the shared representation from which each subject’s decoder is personalized.
  • Communication-aware coordination reduces server-to-client backbone traffic by approximately 42% across five matched realizations.
  • No evaluated strategy is uniformly dominant across datasets and personalization regimes, and cohort averages can conceal subject-level losses.
  • Deployment assessment should jointly consider personalized accuracy, bidirectional communication, update freshness, link-availability sensitivity, and subject-level reliability.
  • The evidence is limited to offline replay, public laboratory datasets, predefined Bernoulli availability profiles, fixed policy parameters, and five matched realizations.

FUNDING

The supplied passages document funding, data provenance, reproducibility materials, and evaluation conventions. They do not identify a substantive funding result beyond institutional support and no specific external grant.

  • Funding: The work received institutional research support associated with the Melchor Visiting Assistant Professorship and the Lucy Family Institute for Data and Society.
  • Funding: No specific external grant supported this work.
  • Data and ethics: The analysis used existing publicly available, de-identified BCICIV-2a and OpenBMI EEG data without recruiting participants or collecting new human-subject data.
  • Reproducibility: The datasets, NEXUS-MI code, experiment outputs, and reproducibility materials are publicly available through repositories, GitHub, and Zenodo.
  • Evaluation conventions: Table S1 averages accuracy equally across k ∈{15, 20, 30} calibration budgets within subjects before computing cohort means.
  • Operating points: P5 and P6 produce the lowest server-to-client traffic because they use stale-aware backbone downloading.

A. Experimental Design and Statistical Analysis

The repeated P3/P5 comparison paired five matched realizations, using hierarchical bootstrap to assess robustness. Communication savings were consistent, while accuracy effects were smaller and realization-dependent.

  • Experimental design: Five matched replicates paired model, availability-trace, and tie-break seeds while keeping P5 priority scheduling fixed.This design varied training initialization and realized gateway availability without changing the scheduling rule.
  • Statistical analysis: The crossed hierarchical bootstrap used 10,000 draws, resampling replicates and subjects while preserving within-subject P3/P5 pairing.This evaluates robustness across matched stochastic realizations.
  • Accuracy robustness: Accuracy effects were realization-dependent: repeated P5-minus-P3 confidence intervals included zero across BCICIV-2a and OpenBMI conditions.BCICIV-2a mean effects were +0.80 and +0.88 percentage points, while OpenBMI effects were −0.09 and −0.07 percentage points across regimes.
  • Subject-level reliability: Subject-level effects included both positive and negative changes, with full subject-by-replicate ranges reaching [−6.94, 10.19] percentage points on BCICIV-2a.Subject-mean ranges were wider under BCICIV-2a EIB-PH than under OpenBMI.
  • Communication outcomes: 42.23% and 41.98% were the mean server-to-client backbone-traffic reductions on BCICIV-2a and OpenBMI, respectively.The reductions ranged from 41.59% to 42.76% on BCICIV-2a and from 41.47% to 42.47% on OpenBMI.
  • Communication outcomes: P5 reduced total traffic but increased rejected-upload rates and accepted-update staleness, while buffered-upload delay changed by no more than 0.003 rounds.Mean total-traffic decreases were 3.20 MB on BCICIV-2a and 15.91 MB on OpenBMI.

S3. LINK-AVAILABILITY SENSITIVITY

The availability sensitivity analysis varied heterogeneous gateway online probabilities while holding P3 and P5 buffering and stale-update admission fixed. P5 consistently reduced server-to-client traffic, but its accuracy advantage was not profile-invariant.

  • Availability profiles: Mild, default, and severe profiles assigned high/moderate/low gateway online probabilities of 0.98/0.85/0.60, 0.95/0.70/0.40, and 0.90/0.50/0.20.P3 and P5 retained the same FIFO buffering and stale-update admission rules across profiles.
  • Synchronization effects: As availability decreased, rejected uploads increased and severe-profile buffered-upload delay reached approximately 1.6 rounds.Accepted-update staleness remained higher under P5.
  • Communication outcomes: P5 used less server-to-client traffic at every severity level, with BCICIV-2a traffic of 4.07, 3.68, and 2.99 MB versus 7.46, 6.27, and 4.86 MB under P3.The sensitivity analysis therefore shows persistent communication reduction as availability severity changes.
  • Accuracy sensitivity: Accuracy contrasts changed direction across profiles: BCICIV-2a effects were +1.77, −0.03, and +0.16 points, while OpenBMI effects were +0.08, +0.99, and −0.46 points.There was no profile-invariant accuracy advantage for P5.
  • Subject-level associations: No recorded communication-exposure quantity showed a consistent association with subject-level accuracy differences across the evaluated datasets and policies.The correlations are exploratory descriptions rather than evidence of causal mechanisms or reliable vulnerability predictors.
Loading 2609.09786v1…