PaperlayerView exact revision on arXiv

Source-linked AI summary

XAI-Refine: An Automated Explanation-Knowledge Loop for Brain-Age Prediction

Yang Qiao, Junjie Wu, Deqiang Qiu, James J. Lah, Liang Zhao

arXiv:2609.09388v1cs.LGq-bio.NC

TL;DR

Accurate brain-age predictions do not by themselves establish reproducible, neurobiologically supported mechanisms, and existing workflows often stop at diagnosis or prespecify correction targets. XAI-Refine iteratively converts reliable explanations and verified literature into typed, minimum-change constraints while preserving explanatory semantics. In functional-connectivity experiments, the final model improves prediction, explanation confirmation, alignment, and target-direction realization with little detected off-target change.

  • Problem

    Accurate predictions and post-hoc explanations do not ensure reproducible, faithful, neurobiologically supported mechanisms or provide an automatically derived correction target.

  • Method

    XAI-Refine consolidates reliable multi-run explanations, verifies neutral literature queries, projects explanations onto knowledge-induced admissible sets, and translates revisions into typed constraints.

  • Results

    The final model reduces MAE from 4.5360 to 4.3813, increases PCC from 0.4788 to 0.5274, raises Confirmation from 0.4000 to 0.6500, and improves AEA from 0.3646 to 0.5722 versus prediction-only f₀.

  • Takeaways & Limitations

    The framework provides a structured route from post-hoc model interpretation to evidence-guided, target-specific refinement with bounded non-target explanatory drift.

  • Takeaways & Limitations

    The study is limited to resting-state functional connectivity, requires evaluation across populations, sites, and imaging protocols, and is not intended for individual clinical decisions.

Abstract

from arXiv · show

Brain-age prediction models are commonly evaluated by predictive accuracy, yet accurate predictions alone do not establish that a model relies on reproducible or neurobiologically supported mechanisms. Post-hoc explanation methods can expose these mechanisms, but existing workflows typically stop at diagnosis or require correction targets to be specified before model analysis. We propose XAI-Refine, an automated explanation-knowledge loop for brain-age prediction from resting-state functional connectivity. At each iteration, XAI-Refine consolidates complementary post-hoc analyses across repeated training runs into reliable, structured model explanations. It converts each reliable explanation into a neutral neurobiological question, retrieves and verifies relevant literature, and compiles the verified evidence into an admissible set in the same typed explanation space. The target for refinement is defined as the minimal projection of the current model explanation onto the admissible set induced by applicable verified knowledge. This revised explanation is then translated into a differentiable constraint while preserving the originating model variable, measurement operator, and applicable scope. Candidate updates are promoted only when multi-seed validation confirms target-directed explanatory movement, predictive performance remains within a prespecified guardrail, and non-target explanatory drift remains bounded. Experiments on functional-connectivity-based brain-age prediction evaluate predictive performance, explanation reliability, literature alignment, and target-specific model revision, illustrating a structured route from post-hoc analysis to evidence-guided model refinement.

1 Introduction

Brain-age prediction from functional connectivity requires more than low error: explanations should be reproducible, faithful, and neurobiologically supported. XAI-Refine addresses this gap by deriving refinement targets from validated explanations and verified evidence.

  • Motivation: Low prediction error does not establish that functional-connectivity models use reproducible or neurobiologically supported mechanisms.Models with similar MAE can differ in prediction slope, age-related bias, and longitudinal consistency.
  • Motivation: Post-hoc explanations can be unstable, unfaithful, or insensitive to model parameters, while stable explanations may still encode shortcuts.Reliable workflows therefore need model faithfulness, empirical reproducibility, and independently verified domain evidence.
  • Gap: Existing explanation-guided methods typically assume the desired prior or correction target is known before model analysis.Related evidence-retrieval systems do not determine how verified claims should minimally revise a specific explanation.
  • Approach: XAI-Refine derives refinement targets by analyzing multiple training runs and explanation families, then representing reliable findings as typed records.The records preserve the model variable, measurement operator, and applicable scope.
  • Approach: Verified evidence induces a minimal explanation revision that is translated into a differentiable constraint on the same variable and operator.Updates are accepted only when explanations move toward the revision without violating predictive guardrails.
  • Evaluation: The framework evaluates prediction, explanation reliability and faithfulness, held-out domain alignment, targeted revision progress, and off-target explanatory drift.These evaluations support a closed prediction–explanation–knowledge loop rather than one-way post-hoc analysis.

2 Related Work

Prior brain-age and explanation-guided learning studies improve prediction or incorporate domain knowledge, but commonly leave interpretation vulnerable to instability and assume correction targets are predefined. XAI-Refine instead discovers candidate relations and constructs scope-aware guidance from verified evidence.

  • Brain-age prediction: Brain-age models use fine-grained connectivity, latent networks, multimodal imaging, and graph architectures, but their interpretations may be unstable, unfaithful, or parameter-insensitive.XAI-Refine retains reproducible and faithful relations, verifies their scientific scope, and uses them for refinement.
  • Explanation-guided learning: Explanation-guided learning constrains gradients or incorporates priors and symbolic or causal knowledge, generally assuming a predefined correction target.Scientific claim-verification systems retrieve and synthesize evidence but do not by themselves define model-specific revisions.

3 Problem Formulation

The paper formulates brain-age prediction as supervised learning from functional-connectivity inputs while requiring explanations to be reliable and consistent with applicable verified knowledge. A structured explanation and knowledge-induced admissible set connect prediction to scientific refinement.

  • Prediction problem: The supervised task uses inputs x_i, prediction targets y_i, auxiliary information m_i, and a predictor f_θ producing predictions.The formulation introduces the prediction problem before adding explanation and knowledge requirements.
  • Prediction problem: In the brain-age application, x_i is a functional-connectivity matrix, y_i is chronological age, and m_i is subject identity.These variables instantiate the general supervised formulation for the target application.
  • Explanation requirements: The structured explanation z_θ retains interpretable patterns that are faithful to the predictor and reproducible across repeated training.This extends the objective beyond accurate prediction to reliable scientific relationships.
  • Knowledge consistency: Each retained explanation becomes a provisional scientific hypothesis that query and verification operators evaluate against a literature corpus.The resulting knowledge record induces a knowledge-consistent explanation set in the structured explanation space.
  • Knowledge consistency: The formulation jointly optimizes predictive loss, structured explainability, and satisfaction of the admissible set induced by verified knowledge.It seeks an evidence-consistent explanation equilibrium connecting predictor-generated hypotheses with verification.

4 XAI-Refine

XAI-Refine iteratively extracts reliable explanations, verifies neutral literature queries, compiles evidence into typed admissible sets, and applies minimum-change updates. The loop preserves explanatory semantics and uses validation to control whether revisions are executable and appropriate.

  • Workflow: The loop maps reliable explanations through query, verification, projection, and predictor-update operators to obtain executable refinements.These stages connect post-hoc analysis, literature evidence, knowledge-consistent revision, and model optimization.
  • Reliable explanations: Each explanation analysis specifies a model quantity, measurement operator, and evaluation scope, which define the explanatory relation being measured.Relations may concern prediction behavior, feature response, longitudinal consistency, attribution, or representations.
  • Reliable explanations: Multiple random-seed runs are consolidated while retaining seed-wise measurements and provenance without changing the analyzed quantity, operator, or scope.This aggregation improves reproducibility while preserving source semantics.
  • Reliable explanations: Reproducibility is assessed through support, coverage, and stability, while faithfulness is tested with procedures such as perturbation or randomization.Only explanations satisfying both validation criteria proceed to literature verification and refinement.
  • Literature verification: Neutral queries preserve the scientific variable, relation type, population, measurement setting, and scope while withholding model-observed direction, magnitude, and reliability.The design reduces confirmation bias during literature retrieval.
  • Literature verification: Verification returns provenance-bearing records containing evidence scope, knowledge-side relations, confidence, status, and supporting publications.The evidence is aligned with the originating model explanation before compilation.
  • Typed revision: A deterministic compiler maps verified relations into typed admissible sets, with stronger evidence producing narrower constraints and absent evidence producing no revision.Unaligned or nonrepresentable evidence is deferred rather than translated into a training constraint.
  • Typed revision: The literature agent does not directly prescribe revision direction or magnitude; projection determines the smallest model-specific change within the admissible set.Executable revisions use type-compatible discrepancy measures while preserving the originating explanatory quantity and operator.

5 Experiments

Experiments evaluate XAI-Refine across prediction, explanation reliability, literature alignment, and targeted revision. Results show that iterative, validated refinement improves the balance between target movement, predictive preservation, and off-target control.

  • Experimental design: The experiments jointly assess prediction, explanation reliability, domain alignment, and controlled explanatory revision across baselines, refinements, and ablations.The evaluation includes subject-disjoint longitudinal data, multiple predictor families, matched refinement variants, and component stress tests.
  • Overall comparison: XAI-Refine achieves the best MAE and PCC among matched refinement strategies while improving Faithfulness, AEA, and target Progress over the prediction-only parent.Direct evidence-to-loss and one-shot revision lead selected individual metrics but do not match the complete loop across prediction, alignment, and progress.
  • Overall comparison: MAE falls from 4.5360 to 4.3813, PCC rises from 0.4788 to 0.5274, Confirmation rises from 0.4000 to 0.6500, and AEA rises from 0.3646 to 0.5722 relative to f_0.The final model realizes 84.44% of prespecified target directions with almost no detected off-target change beyond prediction-only retraining variation.
  • Explanation reliability: Multi-run aggregation improves held-out confirmation and perturbation faithfulness, while full reliability filtering reaches Confirmation 0.6500 and Faithfulness 0.8455.The protocol filters three claims while retaining explanations that are both reproducible and behaviorally valid.
  • Verification and revision validity: Variable, operator, and revision-scope preservation exceed 94%, but strict scope agreement is 69.7% because broad evidence can yield overly specific strengthening constraints.The main audit errors concern translating broad or connection-family evidence into overly specific revisions.
  • Targeted refinement: XAI-Refine provides the strongest overall balance between target improvement, update specificity, predictive preservation, and validation-to-test persistence rather than maximizing raw target movement.Direct evidence-to-loss produces larger and more variable off-target change, while unprobed eligible revisions remain below the complete loop.
  • Case study: Aging-related posterior-cingulate and precuneus connections receive direction-neutral importance revision because the evidence supports relevance but not a reliable signed FC–age relation.This case demonstrates that the revision follows the specificity of verified evidence rather than imposing an unsupported response direction.
  • Component ablation: Ablations show that reliability filtering and scope-aware verification determine eligible revisions, while probing and sparse selection reduce off-target change during realization.Removing reliability filtering reduces confirmation and progress; removing probing or sparse selection preserves raw progress but introduces detectable off-target changes.

6 Limitations and Ethical Considerations

The framework depends on reliable explanations, literature retrieval, and evidence verification, and its scope is limited to resting-state functional connectivity. External evaluation is needed across populations, sites, and imaging protocols, while demographic and dataset biases and nonclinical use constrain interpretation.

  • The framework depends on the reliability of post-hoc explanations, literature retrieval, and evidence verification.
  • Incomplete, conflicting, or mis-scoped evidence may lead to inappropriate abstention or revision.
  • The study is limited to resting-state functional connectivity and requires evaluation across populations, acquisition sites, and imaging protocols.
  • Brain-age predictions and explanations may reflect demographic or dataset biases and are not intended for individual clinical decision-making.

7 Generative AI Usage

Generative AI tools supported manuscript editing, organization, debugging, and implementation, but were not treated as scientific evidence or used for final scientific decisions. The authors reviewed and verified the resulting work.

  • Generative AI tools supported manuscript editing, organization, code debugging, and implementation support.
  • The tools were not treated as sources of scientific evidence or used to make final scientific decisions.
  • The authors reviewed and verified generated text, code, literature claims, citations, analyses, and conclusions.

8 Conclusion

XAI-Refine closes the loop from reliable post-hoc model behavior to literature-verified, minimally revised explanations and validated training constraints. Its brain-age instantiation evaluates predictive performance alongside explanation reliability, domain alignment, targeted revision, and explanatory drift.

  • XAI-Refine turns reliable post-hoc model behavior into literature-verified, minimally revised explanations and validated training constraints.
  • The framework keeps scientific evidence, explanation geometry, and optimization distinct but connected through shared variables, operators, and scopes.
  • The brain-age instantiation evaluates prediction, explanation reliability, held-out domain alignment, target-directed movement, and unintended explanatory drift.
  • The framework provides a testable route from model interpretation to automated scientific model refinement without prespecifying the correction target.

A.1 Multi-seed Post-hoc Analysis Protocol

XAI-Refine builds structured explanations from complementary post-hoc analyses across repeated training runs, evaluating their reproducibility and faithfulness before verification. Reliable explanations are linked to neutral literature questions and refined within the same explanatory dimensions while preserving their originating variables, operators, and scopes.

  • Multi-seed analysis: The protocol applies identical post-hoc analyses across models trained under multiple random seeds, using development data without accessing the test partition.Seed-wise measurements are retained as a model family for cross-run analysis.
  • Aggregation: Measurements are aggregated with relation-specific operators while preserving the analyzed variable, measurement operator, scope, direction, numerical values, and provenance.Subject-balanced aggregation prevents participants with more repeated scans from receiving greater total weight.
  • Analysis dimensions: Complementary analyses characterize prediction behavior, longitudinal consistency, feature response, attribution structure, and representation geometry.These analyses capture distinct explanatory relations, including age-specific bias, repeated-scan behavior, feature dependence, anatomical organization, and latent-space structure.
  • Reliability assessment: Reproducibility requires evaluable cross-seed support, broad measurability, stability in the relevant explanation space, and sign agreement for directional relations.Effect magnitude remains part of the model relation rather than being folded into the reproducibility score.
  • Reliability assessment: Faithfulness validation tests whether each explanatory relation matches actual predictor behavior, using validation operators selected for the relation type.Examples include controlled perturbations for feature responses, randomization tests for attributions, observed-output comparisons for prediction-level relations, and independent remeasurement for representations.
  • Verification and refinement: Only analyses meeting both reproducibility and faithfulness requirements enter verification; failed records remain auditable but generate no literature queries or refinement constraints.Reliable records may be consolidated for presentation only when their scientific entity, relation type, scope, and direction agree, without changing atomic source analyses.
  • Verification and refinement: Neutral literature questions omit model-observed effects to limit confirmation bias, while verified evidence defines an admissible set for minimal, dimension-preserving explanation refinement.The resulting correction remains tied to the same quantity and explanatory dimension as the diagnosed discrepancy.

D Details of Compared Methods

The compared methods span predictive baselines, controlled refinement variants, and metrics for prediction, explanation reliability, literature alignment, and revision behavior.

  • Predictive baselines: Predictive baselines include classical vector models, neural and graph models, and brain-network-specific architectures operating on functional-connectivity features.The brain-network-specific group includes BrainNetCNN, BrainGNN, Brain Network Transformer, recent brain-age GNN and transformer models, and BioBGT.
  • Predictive baselines: All predictive baselines use subject-disjoint partitions, with matched seeds, targets, model-selection data, and evaluation procedures where architectures permit.
  • Controlled refinement variants: Controlled refinement variants share the prediction-only TransformerConv backbone and matched experimental settings while varying guidance, target selection, direction, timing, and revision coverage.Variants include prediction-only, sham, random-target, opposite-direction, direct-evidence-to-loss, one-shot, all-eligible, and complete-loop strategies.
  • Evaluation metrics: The evaluation organizes eight primary metrics across prediction, explanation, domain alignment, and controlled revision dimensions.MAE and PCC assess prediction; Confirmation and Faithfulness assess explanations; CSR and AEA assess domain alignment; Progress and Drift assess revision behavior.
  • Explanation metrics: Explanation claims are generated from repeated discovery runs, with candidate edges selected by mean attribution and frozen when at least four of five seeds agree on direction.Held-out Confirmation requires recurring importance and directional agreement, while Faithfulness tests whether perturbing a claimed edge changes predictions in the claimed direction.
  • Domain-alignment metrics: CSR measures literature coverage among stable high-attribution edges, whereas AEA evaluates whether stronger evidence receives greater attribution mass and higher attribution rank.CSR does not distinguish exact from connection-family evidence or account for attribution magnitude and rank.

E.4 Controlled-revision Metrics

Controlled-revision metrics evaluate whether a revised predictor moves prespecified targets in the intended direction while limiting changes to non-target properties.

  • Metric scope: Progress and Drift are defined only for matched parent-to-revised predictor transitions, not independently trained predictive baselines.
  • Target Progress: Progress evaluates nine revision targets and counts the fraction that improve in their intended direction across seeds.It measures the probability that a randomly selected prespecified target improves rather than averaging heterogeneous change magnitudes.
  • Target Progress: All refinement variants are evaluated against the same nine targets defined by the complete XAI-Refine procedure.
  • Non-target Drift: Drift measures changes in seven non-target dimensions, excluding explicitly targeted edges from attribution-based calculations.The natural retraining threshold is estimated from pairwise distances among independently trained prediction-only models.
  • Non-target Drift: Lower Drift is better; zero means no non-target dimension changes beyond ordinary prediction-only variability, while sham continuation has Progress and Drift equal to zero.
  • Evaluation protocol: Explanation discovery uses validation data, freezes claims before testing, and evaluates them on held-out subjects and independently trained confirmation models.The discovery stage does not access test subjects.

F.1.2 Ablation Protocols.

The ablation protocols compare explanation-discovery procedures and evidence-grounded revision components, then audit whether retrieved evidence supports the proposed scope and constraints.

  • Explanation discovery: Single-seed discovery produces 33 claims without cross-seed agreement or intervention-based faithfulness, while multi-seed discovery produces 32 claims after directional agreement.
  • Explanation discovery: Twenty-nine of 32 multi-seed claims pass the additional intervention-based faithfulness gate.
  • Evidence audit: The audit uses 37 purposively stratified evidence-grounding records containing 572 candidate literature records rather than a random corpus sample.
  • Review protocol: Evidence review hides model claims and proposed revisions, after which revision review assesses admissibility, variable and measurement preservation, scope, and minimality.
  • Audit findings: Reviewer labels comprise nine support, two mixed, 26 insufficient, and zero contradiction records, with abstention recommended for 28 records.Variable preservation is accepted for all 37 records, but indirect evidence is sometimes over-compiled into overly specific constraints.
  • Rejected revisions: Five proposed strengthening revisions are judged inadmissible because their optimization targets are more specific than the supplied literature evidence.The principal scope failure is using network-level or broad anatomical relevance to justify connection-family conclusions.
  • Limitations: The audit is limited by one reviewer, purposive sampling, non-identical taxonomies normalized after review, and Loop-0 records rather than the final Loop-6 checkpoint.

I.5 Ablation Results

Ablations show that explanation reliability, scope-aware verification, minimal revision, variable–operator preservation, response probing, and portfolio selection each support different aspects of refinement. Their effects are measured within the implemented dataset, predictor, verification pipeline, and refinement schedule.

  • Removing reliability filtering reduces held-out explanation Confirmation and target Progress, supporting reproducibility and faithfulness screening before model revision.
  • Removing scope-aware verification preserves relatively high Confirmation but substantially lowers Progress when retrieved evidence does not match the required scope.
  • The non-minimal variant worsens prediction and explanation Confirmation, indicating that stronger knowledge-consistent intervention is not necessarily preferable to the smallest evidence-supported revision.
  • Removing variable–operator preservation reduces Progress, supporting constraints that act on the same explanatory quantity that generated the verified hypothesis.
  • Removing response probing or sparse portfolio selection retains relatively high target Progress but increases off-target Drift, indicating greater unintended changes in non-target behavior.
  • Ablation effects may depend on the dataset, backbone, hyperparameters, and available evidence corpus, so they should not be treated as universal effect sizes.
Loading 2609.09388v1…