2.2 Data s ource and s earch s trategy

Metadata were retrieved from the OpenAlex Works application programming interface (API) on 13 July 2026 for English-language records published from 1 January 2018 through 13 July 2026. Earlier seminal papers were retained separately for conceptual grounding and were not modeled. OpenAlex was the only scholarly metadata source queried; Web of Science, Scopus, IEEE Xplore, ACM Digital Library, PubMed, and other databases are not reported as searched.

Twelve relevance-ranked OpenAlex queries were organized into A AND B, A AND C, and A AND B AND C families, where A represented XAI anchors, B represented methods/frontiers, and C represented decisions or applications. Each query was capped at 350 records; OpenAlex meta-counts and downloaded counts were preserved. The method anchors included SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). The exact query strings were:

(1) explainable artificial intelligence SHAP LIME counterfactual explanation; (2) interpretable machine learning feature attribution rule sets local surrogate models; (3) explainable artificial intelligence knowledge graph causal explanation uncertainty robustness; (4) explainable artificial intelligence large language models foundation models multimodal generative AI; (5) explainable artificial intelligence decision support healthcare finance law education; (6) interpretable machine learning high stakes decision making clinical decision support; (7) explainable AI environmental science cybersecurity agriculture risk management; (8) human centered explainable artificial intelligence stakeholder decision governance; (9) explainable artificial intelligence SHAP decision support healthcare; (10) counterfactual explanation algorithmic recourse decision making; (11) interpretable machine learning causal uncertainty decision support; (12) explainable artificial intelligence foundation model decision support.

2.3 Screening and c orpus c onstruction

The API returned 4,200 records. Deduplication was performed in a fixed sequence: normalized DOI (3,222 retained), exact normalized title (3,169), and high-similarity normalized title at a SequenceMatcher threshold of 0.965 (3,166). Title/abstract/keyword screening required a substantive XAI anchor plus a method/frontier or decision/application focus and retained 694 records. Eligibility required an adequate English abstract, an allowed scholarly record type, and substantive XAI focus; 666 records remained. Explicit exclusions covered particle and high-energy physics, generic prediction or optimization without substantive explanation, editorials, notices, corrupted metadata, and incidental mentions of explainability.

After preprocessing and removal of empty vectors, 666 records entered LDA; seven seminal papers were retained outside the model. Figure 1 reports these counts and distinguishes retrieval, deduplication, screening, eligibility, and modeling.

Figure 1. Literature retrieval and screening process for the LDA-assisted structured review

2.4 Text p reprocessing and d ocument -t erm m atrix

The text modeling combined titles, abstracts, and OpenAlex keywords; broad OpenAlex concept labels were retained only as metadata and were not modeled. Processing included lowercase normalization, removal of URLs and non-letter characters, whitespace tokenization, WordNet noun-then-verb lemmatization, standard and project-specific stopwords, and deterministic phrase mapping for terms such as large_language_model, counterfactual_explanation, algorithmic_recourse, feature_attribution, concept_bottleneck, integrated_gradients, and human_centered. CountVectorizer used min_df=3, max_df=0.55, and keep_n/max_features=2,200. The final document-term matrix contained 666 rows and 2,200 terms.

2.5 Topic- n umber s election and LDA e stimation

Candidate models with k=5–16 were estimated using batch variational Bayes for 15 iterations with seed 42. Default symmetric document-topic and topic-word priors (1/k) were used. Additional seeds 123 and 2026 were used to assess matched-topic stability. Diagnostics included c_v and u_mass coherence, conventional perplexity (lower is better), per-token log likelihood, matched-topic cosine stability, and mean pairwise Jensen–Shannon distance. Coherence metrics followed established evaluation studies [14-17]. The selected k=8 solution achieved c_v=0.5226, u_mass=-1.9445, perplexity=1136.03, stability=0.6809, and mean topic separation=0.5824. Although smaller models had lower perplexity and higher stability, k=8 had the highest weighted selection score and clearer substantive separation.

Figure 2 shows the full coherence and model-fit comparison prior to the selected solution.

Figure 2. Topic-number selection using coherence and model-fit diagnostics

2.6 Topic v isualization, i nterpretation, and m anual c oding

Intertopic geometry was calculated from Jensen–Shannon distances between normalized topic-word distributions and projected using metric multidimensional scaling; this avoids treating raw topic weights as Euclidean principal component analysis (PCA) coordinates. The visualization follows the interpretive intent of LDAvis [18].

Topic interpretation used a single, author-assisted computational coding pass. For each selected topic, the top 30 terms and top 20 documents by posterior probability were inspected; off-topic records were excluded before the final rerun; labels were assigned only after semantic inspection; and two to four representative papers were selected from the verified documents. No independent double screening, inter-rater agreement, or Cohen's kappa is claimed.

Figure 3 presents the final eight-topic distance map; the manually interpreted themes are reported in Table S2 in the Supplementary Material.

Figure 3. Intertopic-distance visualization for the selected eight-topic LDA model

2.7 Integration with the PDR f ramework

The LDA topics are treated as an empirical map rather than a substitute for conceptual synthesis. Method-centered topics inform descriptive accuracy; clinical, governance, security, and domain-use topics inform relevance and predictive risk; and counterfactual themes connect explanation to feasible action. Sections 3–7 therefore use the PDR framework to compare methods, identify stakeholder-dependent explanation needs, and translate topic gaps into a research roadmap.

3. PDR f ramework

3.1 Motivation and d efinition

PDR evaluates an XAI system through three distinct questions: whether the predictor is reliable, whether the explanatory artifact faithfully describes the predictor, and whether that artifact is relevant to a defined audience and decision. Keeping these dimensions separate prevents a persuasive visualization from being mistaken for evidence of model quality and prevents predictive accuracy from being mistaken for decision legitimacy [4, 5, 19].

3.2 Predictive a ccuracy

Predictive accuracy concerns generalization to the target population, not only performance on a convenient test split. Evaluation should include calibration, subgroup errors, distribution shift, and robustness to reasonable perturbations. Explanations built on a poorly calibrated or unstable predictor cannot serve as strong decision evidence, even when the explanatory procedure is internally consistent.

3.3 Descriptive a ccuracy

Descriptive accuracy concerns fidelity to the model behavior that an explanation purports to describe. Relevant evidence includes local surrogate fit, perturbation fidelity, sensitivity, randomization checks, and consistency across functionally similar models [20-21]. The relationship between accuracy and interpretability is context dependent rather than universal. In structured, high-stakes tabular tasks, interpretable models can match or exceed black-box performance and should be tested directly; in image, speech, and large-language tasks, expressive models may retain predictive advantages [6, 22].

3.4 Relevance

Relevance asks whether an explanation improves performance on a real task for an intended stakeholder. A clinician may need calibrated risk estimates and concise case evidence; an applicant may need contestable reasons and feasible avenues for recourse; an auditor may need global constraints, stability tests, and traceable documentation. Human satisfaction alone is insufficient if the explanation increases automation bias or does not improve decision quality.

3.5 Mapping m ethodological c ategories to PDR

Intrinsically interpretable models expose decision structure through coefficients, smooth effects, monotonicity, trees, or rules. Post-hoc explainability methods analyze a trained predictor through attribution, local surrogates, examples, concepts, counterfactuals, semantic paths, or visual evidence. Intrinsic structure can improve auditability but still requires validation; post-hoc methods can be practical for expressive models but risk rationalizing them rather than faithfully explaining them. Figure 4 shows how task context, PDR evidence, method selection, stakeholder use, and monitoring form a single evaluation loop.

Figure 4. Decision-oriented PDR framework for XAI method evaluation

3.6 Operationalization, a ssessment, and r eporting

A PDR report should combine predictive evidence (test, subgroup, calibration, and shift performance), descriptive evidence (fidelity, stability, sensitivity, and model dependence), and relevance evidence (task outcome, comprehension, time, contestability, and downstream harm). The intended user, decision, available actions, and institutional constraints should be specified before selecting the explanation. This ordering turns XAI evaluation from a visualization exercise into a decision-support protocol.

4. Interpretable m ethods

Figure 5 organizes representative approaches into intrinsically interpretable models and post-hoc explainability methods. The categories indicate where interpretability enters the system; they do not guarantee that an output is faithful or useful. Each method must still be evaluated against the PDR requirements of the deployment context.

Figure 5. Decision-oriented taxonomy of XAI methods under the PDR framework

4.1 Intrinsic i nterpretable m odels

4.1.1 Linear and l ogistic r egression

Here is the conditional probability, and each represents the change in the log-odds for a one-unit increase in , holding other features fixed. Logistic coefficients are globally interpretable, but nonlinear relationships, interactions, collinearity, and distribution shift can make a seemingly simple model misleading in practice.

4.1.2 Additive and s parse a dditive m odels

The terms encode selected two-way interactions. Generalized additive models with pairwise interactions (GA2M/GAMI) generally constrain the number and complexity of interactions so that the resulting surfaces remain auditable. Sparse additive models (SpAM) impose sparsity across component functions [23-24]. These models often provide a strong PDR compromise for tabular decisions because global effects can be audited without assuming that every relationship is linear.

4.1.3 Decision t rees, r ule l ists, and r ule s ets

Ordered rule lists apply rules sequentially, so each decision depends on earlier else conditions [27]. Unordered rule sets, such as Repeated Incremental Pruning to Produce Error Reduction (RIPPER), CN2, and Interpretable Decision Sets, allow independent rules to fire and can represent multiple sufficient pathways [28-30]. Their main risks are overlap, conflicting coverage, and a rapid loss of simplicity as the rule count grows.

4.2 Post-hoc e xplainability m ethods

4.2.1 Prototypes, c riticisms, and i nfluence f unctions

Here,L is the loss, is the fitted parameter vector, and H_θ̂ is the Hessian of the empirical risk at the solution. This approximation assumes differentiability and a well-conditioned Hessian. However, non-convexity, damping choices, and inverse-Hessian approximation can affect rankings [32].

4.2.2 SHAP and f eature a ttribution

where, is the complete feature set, is a coalition excluding , is the number of features, and is the value assigned to coalition under a specified conditional or interventional semantics. While SHAP offers axiomatic consistency, factors such as dependence-sensitive semantics, background data, approximation, and feature representation can materially change the result [33]. Therefore, stable attributions require explicit conditioning assumptions and sensitivity analysis.

4.2.3 LIME and l ocal s urrogate m odels

where is the black-box predictor, is the class of interpretable models, and weights samples by their proximity to . The term L measures local disagreement, while penalizes complexity. Although LIME is model-agnostic and easy to interpret, factors such as neighborhood definition, sampling distribution, representation, kernel width, and random seed can lead to varying explanations [34]. Thus, model agnosticism comes at the cost of sensitivity to neighborhood design.

4.2.4 Counterfactual e xplanations and a lgorithmic r ecourse

where penalizes departure from the original , measures the loss relative to the desired outcome , and λbalances proximity with target attainment. The set A represents feasibility and actionability constraints. A counterfactual is not actionable if it alters immutable attributes, violates causal dependencies, or shifts costs to the affected person. Recourse must therefore distinguish predictive contrast from feasible intervention [35-37].

4.2.5 Knowledge- g raph- b ased e xplanations

Here, and are the head and tail entity embeddings, respectively, and is the relation embedding. Higher scores indicate greater compatibility. Paths and subgraphs can support semantic explanations, yet embedding proximity alone does not constitute a human-readable explanation. Instead, relevance is determined by relation quality, ontology coverage, and path selection [38].

4.2.6 Gradient, p ropagation, and v isual e xplanations

where denotes the model output, x' represents the baseline, α serves as the path interpolation parameter, and i indexes the input coordinate. However, such attributions heavily depend on the choice of baseline and path [39]. Similarly, while gradient-weighted class activation mapping (Grad-CAM) localizes class-relevant convolutional regions [3], visual plausibility should not be confused with model faithfulness. Indeed, randomization and parameter-sensitivity tests demonstrate that some visually compelling saliency maps exhibit surprisingly weak dependency on the learned model weights [21].

4.2.7 Concept- b ased e xplanations

Concept activation vectors and concept bottleneck models explain predictions using human-named concepts rather than raw features [40-41]. Their relevance is particularly high when domain experts reason in terms of concepts, but validity depends on concept definition, annotation quality, completeness, and stability across populations. A concept layer can still conceal measurement bias or omit important latent factors.

4.3 Critical c omparison under PDR

No method dominates all PDR dimensions. SHAP trades axiomatic consistency against dependence-sensitive semantics; LIME trades model independence against sampling instability; counterfactuals trade actionability against causal infeasibility; concept methods trade human relevance against concept validity; saliency maps trade visual accessibility against uncertain faithfulness; intrinsic models trade auditability against possible loss of flexibility in complex unstructured tasks; and post-hoc methods can supply practical evidence while also rationalizing an opaque predictor. Adversarial constructions can deliberately produce reassuring LIME or SHAP explanations for biased models [42]. These tensions are summarized in the method-selection matrix in Table S3 of the Supplementary Material.

5. XAI a pplications a cross d isciplines

Applications are compared using a common decision template: decision problem, data and predictor, explanation family, target user, PDR priority, evaluation evidence, and practical value. This structure distinguishes the presence of an XAI output from evidence that it improves a decision.

5.1 Environmental s cience

Environmental decisions combine spatiotemporal prediction with physical plausibility, public accountability, and transfer across locations. XAI has been used to interpret green-energy labor markets and energy-system models [43-44]. Attribution may identify drivers of load, pollution, or ecological risk, but an explanation that changes under spatial resampling or conflicts with known mechanisms should not guide policy. The dominant PDR priorities are predictive reliability and relevance; scientists and policymakers generally need stable global effects, uncertainty-aware SHAP/GAM summaries, and spatial diagnostics because decisions must transfer beyond a single fitted sample.

5.2 Education

Educational XAI supports early warning, feedback, personalization, and institutional intervention [45-46]. Students need actionable feedback, teachers need diagnostic evidence, and administrators need fairness and subgroup monitoring. Feature importance can stigmatize learners when correlations are treated as causes, and explanations may expose sensitive behavioral data. Relevance is therefore dominant; teachers and learners generally need sparse rule/GAM summaries and constrained counterfactual feedback because explanations must lead to pedagogically feasible actions.

5.3 Cybersecurity

Cybersecurity applications include malware, intrusion, anomaly, and threat detection [47-49]. Explanations must reduce investigation time, prioritize alerts, expose false positives, and resist adversarial manipulation; a static explanation can become obsolete as attackers adapt. Predictive accuracy and analyst relevance are co-dominant; security analysts generally need concise local attributions plus rule or prototype summaries because rapid triage must be paired with stability and adversarial stress testing.

5.4 Finance

Finance uses XAI in credit, fraud, portfolio, and risk decisions [50-52]. The same outputs often must support model validation, regulatory review, applicant communication, and recourse. SHAP can provide consistent additive summaries, but correlated variables and background choices can change their meaning; counterfactuals are useful only when actions are feasible and lawful. Descriptive accuracy and relevance are paramount; auditors and applicants generally need monotonic scorecards or GAMs, stable SHAP analyses with dependence checks, and constrained recourse because decisions must be contestable and auditable.

5.5 Law

In legal settings, XAI supports contestability, procedural fairness, evidentiary scrutiny, and allocation of responsibility [53-55]. A technically faithful feature weight is not automatically a legally sufficient reason; explanations must connect model evidence to admissible factors, institutional procedures, and affected rights. Relevance and descriptive accuracy are paramount; judges, lawyers, regulators, and affected parties generally need traceable rule paths, source-grounded evidence, and constrained local explanations because reasons must support challenge rather than merely persuasion.

5.6 Agriculture

Agricultural XAI supports yield, irrigation, disease, resource-allocation, and sustainability decisions under variable soil, climate, and management conditions [56]. Global feature effects can reveal agronomic drivers, image explanations can localize disease evidence, and counterfactuals can suggest management changes, but transfer across farms and seasons is a central limitation. Relevance and predictive reliability are dominant; farmers and extension agents generally need compact rules or GAM effects, localized visual evidence, and calibrated uncertainty because recommendations must be feasible under local conditions.

5.7 Healthcare

Healthcare XAI spans diagnosis, triage, prognosis, treatment, imaging, and patient communication [57-59]. SHAP and LIME can clarify tabular risk, Grad-CAM can localize imaging evidence, and counterfactuals can frame possible interventions, but none proves clinical validity. Explanations require external validation, uncertainty characterization, subgroup evidence, and workflow testing; otherwise, they can create false reassurance [60-61]. Predictive reliability and relevance are dominant; clinicians and patients generally need calibrated intrinsic models where feasible, clinically validated attribution or localization, and uncertainty-aware summaries because explanations must support professional judgment rather than replace it.

The complete cross-domain comparison is provided in Table S4 in the Supplementary Material.

6. Emerging f rontiers in d ecision- o riented XAI

6.1 Causal, u ncertainty- a ware, and r obust e xplanations

Counterfactual explanations become decision-relevant only when predictive contrast is connected to a feasible intervention. Causal recourse must model immutable variables, dependencies among actions, uncertainty in structural assumptions, and the distribution of burden across groups [37]. Robustness requires testing explanation sensitivity across perturbations, random seeds, model specifications, and distributional shifts. Uncertainty-aware XAI should report confidence in both prediction and explanatory evidence rather than displaying a single deterministic ranking.

6.2 Concept- b ased e xplanations

Concept-based methods can align explanations with clinical signs, legal categories, object parts, or scientific constructs [40-41]. The research frontier is not simply discovering named concepts; rather, it is validating whether concepts are complete, causally meaningful, stable across populations, and actionable under intervention. Concept completeness and leakage tests should accompany human ratings.

6.3 Generative and m ultimodal AI

Multimodal systems combine text, images, signals, tables, and retrieved knowledge. Explanations must identify which modality provides decisive evidence, whether cross-modal evidence is consistent, and whether a generated rationale is grounded in the input. Additionally, generative-AI explainability raises verifiability, interaction, security, and computational-cost requirements [11]. PDR separates output quality from faithfulness and from the user's ability to verify or contest the output.

6.4 Large l anguage m odels and f oundation m odels

Foundation-model behavior reflects an interplay of pretraining data, instruction tuning, retrieval-augmented generation, tools, and context. A fluent chain-of-thought rationale may serve as effective communication while remaining an unfaithful account of the underlying mechanism producing the answer [10, 62]. Consequently, high-stakes applications should pair rationales with source provenance, counterfactual behavior tests, uncertainty communication, retrieval audits, and task-specific human validation. Crucially, mechanistic evidence, behavioral evidence, and user-facing explanation should be treated and reported as different objects.

7. Conclusion and f uture w ork

7.1 Conclusion

This LDA-assisted structured review mapped 666 verified OpenAlex records into eight empirically derived themes and interpreted them through an adopted PDR lens. The synthesis shows that explanation quality is conditional: predictive reliability determines whether the model deserves attention, descriptive accuracy determines whether an explanation reflects model behavior, and relevance determines whether the evidence helps a defined stakeholder decide, contest, audit, or act.

Across domains, method preference follows decision structure rather than popularity. Intrinsic models are especially valuable when global auditability and structured data make them competitive, while post-hoc methods are often necessary for high-dimensional predictors but require explicit fidelity and stability tests. The central practical lesson is to choose the explanation only after defining the user, action, risk, and evidence requirements.

7.2 Limitations and f uture- r esearch r oadmap

This review has bounded limitations. OpenAlex coverage differs from proprietary and domain-specific databases; abstracts and keywords are incomplete for some records; English-language screening introduces language bias; and the 2018–2026 window excludes earlier non-seminal work. LDA relies on bag-of-words assumptions and is sensitive to preprocessing choices, vocabulary thresholds, random seeds, and k. Topic labeling is interpretive, and metadata-based mapping cannot replace full-text synthesis. The rapidly evolving XAI literature means that foundation-model coverage will require periodic updates.

Future work should shift focus from explanation availability to decision evidence. Priorities include standardized PDR-aligned evaluation; context-dependent tests of prediction and interpretability; human-centered task validation; causal and actionable recourse; robustness, uncertainty, drift, and lifecycle monitoring; defenses against explanation gaming; privacy–transparency trade-offs; regulatory and legal accountability; protection of vulnerable stakeholders; multimodal and generative-AI faithfulness; computational and carbon costs; and reproducible deployment governance [19, 42, 63]. Table S5 in the Supplementary Material translates these frontiers into testable questions, recommended evaluation strategies, expected decision value, and responsible stakeholders.

Supplementary information

Supplementary materials associated with this article are available online and can be accessed via the journal website:

Authors' contribution

Jiangshan Zhu: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Data curation, Visualization, Writing—original draft, and Writing—review and editing.

Conflict s of i nterest

The author declares no conflicts of interest.

Funding

This research received no external funding.