Blockchain-Integrated XAI Framework (BXHF)

• The paper introduces BXHF, a framework that combines blockchain for immutable data logging with XAI to produce verifiable, interpretable AI outputs.
• BXHF is defined by dual trust: data-level trust via cryptographic ledgers and decision-level trust through human-digestible explanations, ensuring robust system transparency.
• Practical instantiations in healthcare, finance, and aviation show BXHF enabling secure, transparent, and efficient decision processes across distributed systems.

Searching arXiv for the cited BXHF-related papers and adjacent work to ground the article in current literature. arXiv search query: "Blockchain-Integrated Explainable AI Framework BXHF healthcare (Mohsin, 18 Sep 2025)" Blockchain-Integrated Explainable AI Framework (BXHF) denotes an architectural pattern in which blockchain-mediated integrity, provenance, and access control are coupled with explainable AI so that data exchange is verifiable and model outputs are transparent, auditable, and operationally usable. In its explicit healthcare formulation, BXHF addresses both safe data exchange and comprehensible AI-driven clinical decision-making by combining encrypted patient records, smart-contract-governed access, predictive modeling, formalized interpretability, and hybrid edge--cloud deployment with federated computation across institutions (Mohsin, 18 Sep 2025). Related work generalizes the same design objective beyond healthcare, presenting blockchain as an immutable record-keeping substrate for AI inputs, outputs, model states, explanations, and audit trails in high-stakes settings (Akther et al., 7 Mar 2025).

1. Definition and conceptual scope

BXHF is defined by the conjunction of two trust requirements that are treated as technically distinct but operationally linked. The first is data-level trust, implemented through immutable ledgers, cryptographic hashing, provenance tracking, consent management, and tamper-evident audit trails. The second is decision-level trust, implemented through XAI mechanisms that yield feature attributions, saliency, rules, confidence scores, or other human-digestible rationales for AI outputs (Mohsin, 18 Sep 2025). This dual framing is central to the literature: blockchain is used to secure and attest the lifecycle of data and decisions, whereas explainability methods provide the semantic content that humans inspect.

The motivating problem is the persistence of the AI “black box” in domains where legal, ethical, and operational stakes are high. A survey of blockchain as a platform for AI transparency identifies healthcare, finance, and autonomous systems as archetypal cases in which opaque AI decisions limit understanding, trust, and verification, and where immutable logging can improve decision traceability, data provenance, and model accountability (Akther et al., 7 Mar 2025). In aviation, a related formulation emphasizes fragmented and siloed Air Traffic Management data, conflicts over provenance and access, and the need for secure data chains as autonomy increases; explainable AI is then positioned not merely as a user-interface add-on, but as a mechanism for building stakeholder trust and informing security requirements in the underlying blockchain-network stack (Axon et al., 2023).

The concept therefore refers less to a single canonical implementation than to a family of architectures in which explainability artifacts and blockchain records are bound together. This suggests that BXHF is best understood as a systems pattern for trustworthy AI rather than as a single model class or consensus protocol.

2. Layered architectures and system organization

In the healthcare paper that explicitly names BXHF, the framework is structured as five interdependent layers: a Data Layer holding encrypted patient records $D = \{(x_i, y_i)\}_{i=1}^n$; a Blockchain Layer with a distributed ledger $L$ and smart contracts $\phi$ for access control; an AI Layer for predictive modeling; an XAI Layer for formalized interpretability; and a Deployment Layer implementing a hybrid edge--cloud configuration with federated computation and no central authority (Mohsin, 18 Sep 2025). Sensitive data and inference remain on hospital-owned edge devices, while collaborative model training occurs on federated cloud nodes; blockchain synchronizes trust, consistency, and auditability across sites.

Closely related systems instantiate the same logic with domain-specific layering. IyaCare is described as a modular, service-oriented system with six layers: Presentation, Application, AI Analytics, IoT Integration, Blockchain, and Data. Its stack combines a Next.js/TypeScript/PWA frontend, Firebase services, Ethereum smart contracts, and XGBoost-driven risk prediction, with offline capabilities and SMS-based communication for community health workers (Ankeli et al., 8 Dec 2025). Although IyaCare is framed as an integrated AI-IoT-blockchain maternal-health platform rather than as a BXHF by name, its architecture directly operationalizes explainability, auditability, and secure record handling in the sense used by BXHF.

In financial decision systems, the same decomposition appears with different modules. DeFi TrustBoost uses a web application for knowledge elicitation and consent management, an API server, a split on-chain/off-chain data storage layer, and a human-in-the-loop XAI decision-making system built around a 1D-CNN, with Ethereum or Hyperledger providing tamper-proof logging and consent management (Sachan et al., 28 Nov 2025). In affective AI, “Immutable Explainability” replaces post-hoc explainers with a white-box fuzzy inference engine and couples it to a blockchain anchoring layer so that each explainability artifact can be independently verified after the fact (Fransoy et al., 11 Dec 2025). Across these systems, the recurrent architectural invariant is separation between heavy computation or sensitive content off-chain and cryptographic attestation on-chain.

3. Mathematical and algorithmic formulation

The healthcare BXHF paper gives the clearest formalization of the framework. Access control is represented through a smart-contract function

$$\phi(u, d) = \begin{cases} 1, & \text{if user } u \text{ is authorized to access data } d \ 0, & \text{otherwise} \end{cases}$$

so that permissions become executable policy rather than merely administrative convention (Mohsin, 18 Sep 2025). Predictive modeling is expressed as a function $f: X \to Y$, and training is posed as

$$f^* = \arg\min_{f \in \mathcal{F}} \mathbb{E}_{(x,y)\sim D}[\ell(f(x), y)].$$

Explainability is formalized as an explanation function $g(x, f(x)) \to E$, where $E$ is a human-digestible object such as feature attributions, saliency, or rules, subject to the constraint

$\sum_j \alpha_j = f(x).$

The paper’s distinctive move is to treat interpretability as part of training rather than only post hoc, by adding an explanation plausibility penalty $\Omega(g(f(x)))$ weighted by $L$0 (Mohsin, 18 Sep 2025).

The same source further describes a joint objective function that unifies predictive loss, interpretability, and security. In prose, the optimization targets models that are simultaneously accurate, explanation-compatible, and aligned with a blockchain-derived security metric $L$1, with hyperparameters $L$2 and $L$3 controlling the relative emphasis on interpretability and security (Mohsin, 18 Sep 2025). This is a defining feature of the BXHF formulation: security and explanation are treated as first-class optimization constraints rather than external compliance checks.

Related domains instantiate analogous formal models. In ANFB-AI for FinTech, the digital ecosystem is written as

$L$4

with users $L$5, transactions $L$6, blockchain nodes $L$7, and distributed ledger $L$8. A block is represented as

$L$9

with linkage through

$\phi$0

and validation via

$\phi$1

Threat detection is posed as a classification map $\phi$2, while the decision policy is explicitly three-level: $\phi$3 (Mishra et al., 25 Mar 2026). This formalism does not define BXHF directly, but it shows how blockchain-attested explainability can be embedded in a risk-based decision pipeline outside healthcare.

IyaCare provides an applied model equation rather than a framework-level optimization. Its XGBoost classifier can be summarized as

$\phi$4

where predictions are summed over $\phi$5 regression trees for input features $\phi$6 (Ankeli et al., 8 Dec 2025). In that system, explainability is realized operationally through risk scores, confidence levels, and interpretable recommendations rather than through a standalone formal XAI objective.

4. Explainability, provenance, and auditability mechanisms

A BXHF links explanations to records in ways that are cryptographically attestable. In the healthcare formulation, every data access and write is logged immutably on a distributed ledger, and both predictions and explanations are hashed and recorded so that other participants can verify input-data integrity, correct model invocation, and faithful explanations (Mohsin, 18 Sep 2025). The same paper emphasizes that explanation integrity is a separate object of protection: a cryptographic hash of each explanation is on-chain, so subsequent alteration is detectable by any participant.

This distinction between producing explanations and attesting them is developed further in the broader transparency literature. Blockchain-based AI transparency work describes a generic block containing timestamp, input hash, model hash, output, and explanation, thereby binding a particular decision to a particular model state and input representation (Akther et al., 7 Mar 2025). DeFi TrustBoost makes the tamper test explicit. For an explanation $\phi$7, the off-chain system stores the encrypted artifact, while the blockchain stores $\phi$8. Auditing recomputes the explanation hash after decryption, with tampering defined by

$\phi$9

The same logic is applied to consent-state auditing across organizations (Sachan et al., 28 Nov 2025). This is not merely logging; it is a verification protocol.

Explainability content itself varies by modality and model. The healthcare BXHF paper lists SHAP, LIME, attention-based maps, feature attributions, and rule extraction as admissible explanation methods (Mohsin, 18 Sep 2025). IyaCare operationalizes explainability through confidence scores, interpretable recommendations, visual risk indicators, risk trajectories, alert summaries, surfaced input features, and the reasoning behind high-risk classifications, all designed for healthcare worker comprehension (Ankeli et al., 8 Dec 2025). “Immutable Explainability” uses a white-box Mamdani-type multimodal fusion engine in which rule activations and output sets are preserved as part of the explainability artifact. Its final fusion is

$$\phi(u, d) = \begin{cases} 1, & \text{if user } u \text{ is authorized to access data } d \ 0, & \text{otherwise} \end{cases}$$0

and a multimodal coherence index is logged as

$$\phi(u, d) = \begin{cases} 1, & \text{if user } u \text{ is authorized to access data } d \ 0, & \text{otherwise} \end{cases}$$1

(Fransoy et al., 11 Dec 2025). Here the explanation is not post-hoc feature ranking but an explicit trace of fuzzy inference.

Verifiable execution can also be decentralized. BRAIN uses VRF-based cryptographic sortition to select inference and training committees, a two-phase commit-and-reveal pipeline, smart-contract finalization, and on-chain rewards and slashing. All requests, committee actions, outputs, and consensus outcomes are public and traceable, and the same mechanism is identified as compatible with future BXHF deployments in which explanations, not only raw outputs, are committed, revealed, and attested (Park et al., 2023). A recurrent misconception is that blockchain alone solves the black-box problem. The literature instead separates immutable provenance from explanation generation: the former ensures non-repudiation, while the latter determines whether a human can understand the decision.

5. Domain instantiations and operational use cases

In healthcare, BXHF is presented as applicable to cross-border clinical research networks, uncommon illness detection, and high-risk intervention decision support. The framework supports federated learning across institutions without transferring raw data, while maintaining auditability from data ingestion to clinician-facing recommendation (Mohsin, 18 Sep 2025). IyaCare gives a concrete maternal-health instantiation: a proof-of-concept integrated platform for predictive risk assessment, continuous vital-sign monitoring, and secure health-record management in resource-constrained settings. Its XGBoost classifier trained on the UCI Maternal Health Risk dataset achieved 85.2% accuracy and 85.1% F1, IoT streaming data enabled real-time risk recalculation with <5s latency, and the Ethereum testnet implementation reported 99.8% data integrity with 3.2s transaction latency (Ankeli et al., 8 Dec 2025). The same system logs AI-generated risk assessments and alerts on-chain and supports offline-first operation with SMS-based communication.

In finance, ANFB-AI extends the blockchain-XAI combination to secure transaction processing under uncertainty and behavioural drift. It combines a permissioned blockchain, smart contracts, and an adaptive neuro-fuzzy model whose linguistically interpretable rules make the rationale for risk classification reviewable by regulators and auditors (Mishra et al., 25 Mar 2026). Simulations under normal, medium, and high-fraud conditions report 99.0 accuracy in the normal scenario, 98.8 accuracy in the high-fraud scenario, 96–97 precision in the high-fraud scenario, 0.7–1.2 seconds transaction confirmation time, 70–120 ms block propagation delay, and ~0.8–1.2 seconds end-to-end latency. DeFi TrustBoost focuses more specifically on underserved lending and regulatory auditability. It combines a human-in-the-loop explainable 1D-CNN with on-chain hashes of explanations, consent states, and model configurations, while keeping full explanations, model parameters, and identity data off-chain in encrypted form (Sachan et al., 28 Nov 2025).

In aviation, the relevant architecture is framed as a secure network for Unmanned Traffic Management. Data from airlines, airports, ANSPs, regulators, and UAS providers are exchanged over a permissioned blockchain, while federated CNN/RNN models use air traffic and weather data for trajectory and delay prediction (Axon et al., 2023). Explainable AI methods such as SHAP and LIME expose which data features drive model outputs, and these trust signals are used to backpropagate security requirements to the blockchain controller. Consensus choice is task-dependent: collision avoidance is associated with Byzantine Fault Tolerance and 100 ms latency, congestion management with Raft or PBFT and 100 ms–1 min, delay prediction with proof-based schemes such as PoA or PoS and 1 min, autonomous landing with BFT and ≈10 ms, and navigation planning with Crash Fault Tolerance over seconds.

For large-scale model execution, BRAIN addresses a different but adjacent concern: trustworthy decentralized inference and training when models are too large for naive blockchain execution. Its randomly selected committees, pipelined commit-and-reveal protocol, and smart-contract coordination yielded tasks-per-second performance 454.4293 times greater than that of a naive single-phase implementation (Park et al., 2023). While BRAIN is not itself an XAI framework, it provides an execution substrate that is directly relevant when a BXHF must attest not only explanations but also the correctness and provenance of large-model inference and training.

In affective AI, “Immutable Explainability” validates a white-box plus blockchain pattern on the Spanish MEACorpus 2023. The system anchors only the SHA-256 hash of each redacted explainability artifact on Sepolia Testnet, keeping the full artifact encrypted and off-chain under user control. Reported latency is ~3.9s for ASR, <0.05s for fusion, and ~4s total, with blockchain anchoring asynchronous and adding no user-facing latency; the fuzzy-fusion model achieved weighted F1 0.429 compared with 0.408 for linear fusion and 0.396 for text-only (Fransoy et al., 11 Dec 2025). This instantiation is especially relevant to BXHF because it treats the explanatory trace itself as the primary object to be made immutable.

6. Limitations, design tensions, and open research directions

The literature consistently identifies scalability, integration complexity, and computational overhead as unresolved constraints. A general survey on blockchain-enhanced AI transparency notes that low transaction rates, especially in Proof-of-Work systems, can bottleneck real-time or high-frequency AI pipelines, while the combination of blockchain and AI increases design and operational complexity (Akther et al., 7 Mar 2025). IyaCare makes the same point more concretely by observing that Ethereum’s native cost and throughput are a challenge for scale and suggesting hybrid approaches, low-fee blockchains such as Polygon, or permissioned chains such as Hyperledger (Ankeli et al., 8 Dec 2025). “Immutable Explainability” similarly points to Layer-2 solutions and Merkle aggregation for scalability (Fransoy et al., 11 Dec 2025).

A second tension is between immutability and privacy regulation. DeFi TrustBoost states the issue directly: immutable public ledgers conflict with deletion or redaction requirements such as the “Right to be Forgotten,” and blockchain storage is expensive for large artifacts (Sachan et al., 28 Nov 2025). The dominant response across the papers is therefore an on-chain/off-chain split: only hashes, pointers, consent states, or minimal metadata are placed on-chain, while sensitive data, full explanations, or model parameters remain encrypted off-chain (Sachan et al., 28 Nov 2025). The healthcare BXHF paper extends the same logic through homomorphic encryption and privacy-preserving query protocols, while keeping federated training parameter exchange separate from raw-data movement (Mohsin, 18 Sep 2025).

A third limitation concerns evidential maturity. IyaCare explicitly reports reliance on synthetic validation data and simulated healthcare environments (Ankeli et al., 8 Dec 2025). The blockchain-transparency survey observes a lack of widespread real-world validation (Akther et al., 7 Mar 2025). “Immutable Explainability” describes its current pipeline as heuristic, corpus-specific, and not expected to be state-of-the-art in accuracy, with no user studies yet and no fully implemented SSI/VC stack (Fransoy et al., 11 Dec 2025). In aviation, the secure explainable blockchain ecosystem for UTM is characterized as an emerging area requiring significant research and development by the community (Axon et al., 2023).

The research frontier therefore lies less in proving that blockchain and XAI can be combined than in stabilizing the combination under realistic deployment constraints. Proposed directions in the cited work include zero-knowledge proofs, permissioned chains with low-latency consensus, full SSI integration through decentralized identifiers and verifiable credentials, state-of-the-art neural modules plugged into auditable pipelines, cognitive networking for adaptive consensus and resource allocation, and broader empirical validation in live settings (Axon et al., 2023). This suggests that BXHF is evolving toward infrastructures in which explanation fidelity, cryptographic attestation, privacy preservation, and distributed systems performance must be jointly optimized rather than handled as separable engineering layers.