Journal of Primeasia

1. Introduction

Artificial intelligence has, in just over a decade, moved from a peripheral curiosity in hospital corridors to something resembling infrastructure — quietly shaping how clinicians read scans, flag risk, and triage patients (Pinto-Coelho, 2023). Clinical decision support systems built on AI promise to synthesize sprawling, heterogeneous data — electronic health records, laboratory panels, imaging studies — into something a busy physician can actually use in the few minutes available between patients. That promise is real. It is also, frustratingly, still mostly unrealized at scale, and the reasons are not especially mysterious so much as structural.

Chief among them is fragmentation. Patient data rarely lives in one place; it is scattered across hospitals and health systems that seldom communicate with one another, each holding only a partial picture (Amjad et al., 2023). A model trained on a single hospital's population tends to absorb that hospital's quirks — its demographics, its equipment, its documentation habits — and generalizes poorly elsewhere. The obvious remedy, pooling data into a central repository, runs almost immediately into a second, harder problem: privacy law. HIPAA, GDPR, and the various institutional policies layered on top of them exist precisely to prevent the kind of casual data movement that centralized training would require (Shahid et al., 2022). Hospitals are, understandably, reluctant to hand over raw records, and the legal and logistical overhead of attempting to do so tends to stall multi-institutional projects before they get anywhere (Bharati & Podder, 2022; Dicuonzo et al., 2023).

Federated learning offers a way around this impasse — though "around" may be too tidy a word, since the approach trades one set of difficulties for another. Rather than moving data to a central model, FL moves a model to the data: each hospital trains locally, and only the resulting parameters, not the underlying records, are shared with a central aggregator (Kumar & Singla, 2021; Brisimi et al., 2018). In principle, this allows institutions to benefit from each other's data diversity without ever exposing an individual patient record. Several authors have described FL as something close to a regulatory and ethical sweet spot for healthcare AI (Rahman et al., 2023). Whether it lives up to that description in practice is a separate question, and one this review tries to engage with rather than simply assume.

A second, somewhat less obvious obstacle sits underneath the privacy question: a federated model that performs well is not automatically a model clinicians will trust. The architectures best suited to clinical prediction — deep neural networks, in particular — are notoriously opaque, producing outputs without producing reasons. That opacity is not a cosmetic flaw; it is a genuine barrier to adoption (Buhrmester et al., 2021). A clinician asked to act on a prediction they cannot interrogate will, not unreasonably, hesitate to do so. Explainable AI methods — SHAP, LIME, attention-based visualizations among them — offer a partial answer (Abdullah et al., 2021; Y.-P. Zhang et al., 2023).

What remains comparatively underexplored is the combination of all three demands at once: training that respects institutional data boundaries, predictions clinicians can actually interrogate, and privacy guarantees robust enough to satisfy regulatory scrutiny, without any one requirement quietly undermining the other two (Ogrezeanu et al., 2022). This review surveys the literature across these three domains — federated training architectures, privacy-preserving mechanisms, and explainability methods — to identify recurring patterns, persistent gaps, and the points where these strands have (and have not) been brought together. Building on that synthesis, the final sections propose PPFBXAIO, a conceptual framework intended to integrate these elements into a coherent design for multi-hospital clinical decision support — offered here as a direction for future empirical work rather than a tested solution (Mylrea & Robinson, 2023; J. Zhang & Zhang, 2023).

2. Review Methodology

This review adopts a narrative rather than systematic format, reflecting its purpose: to synthesize conceptual and methodological patterns across a fragmented and fast-moving body of literature, rather than to exhaustively catalogue every study touching federated learning in healthcare. Sources were identified primarily through targeted searches of PubMed, IEEE Xplore, Scopus, and Google Scholar, using combinations of terms including "federated learning," "explainable AI," "privacy-preserving machine learning," "differential privacy," "homomorphic encryption," "blockchain," and "clinical decision support," cross-referenced with "hospital," "healthcare," and relevant clinical domains such as cardiology and oncology. Reference lists of identified articles were also examined to locate additional sources — a snowball approach reasonably common in narrative reviews of technically fast-moving fields.

Priority was given to peer-reviewed journal articles and conference papers published between 2018 and 2023, reflecting the period during which FL and XAI methods matured sufficiently for healthcare-specific application; a small number of earlier sources are retained where they provide definitional or historical grounding. Because this is a narrative rather than systematic review, no PRISMA flow diagram or formal inter-rater screening process was applied — inclusion was guided by topical relevance and methodological clarity, and this limitation is reported openly rather than implied otherwise.

Studies whose primary focus fell outside healthcare or clinical informatics — even where superficially adjacent to FL or XAI terminology — were excluded from the evidentiary synthesis in Table 1 unless their methodological contribution was directly and explicitly transferable to a clinical use case. This criterion was applied retrospectively during preparation of this review to remove several citations that, on closer reading, did not support the claims originally attributed to them; those corrections are noted in Section 7.

3. Federated Learning Approaches in Multi-Hospital Healthcare Settings

Federated learning's appeal in healthcare rests on a fairly simple inversion of the usual machine learning workflow: instead of bringing data to the model, FL brings the model to the data. Brisimi et al. (2018) demonstrated one of the earlier applications of this idea to electronic health records, showing that predictive models trained federated across institutions could approach the performance of centrally trained equivalents — an important early proof of concept, since it suggested the privacy benefit need not come at a steep accuracy cost. Subsequent work has extended this logic across a fairly wide range of clinical tasks. Bebortta et al. (2023), for instance, proposed FedEHR, a federated framework specifically targeting heart disease prediction from IoT-connected electronic health records, while Wang et al. (2023) applied federated domain translation to synthesize cross-modality brain imaging data without centralizing the underlying scans — a somewhat different application of the same underlying principle.

What's striking, reading across this literature, is how consistently FL is framed not as a performance optimization but as an access-enabling technology — a way of making multi-institutional collaboration possible in contexts where it otherwise wouldn't be, given regulatory constraints (Rahman et al., 2023; Kumar & Singla, 2021). Sandhu et al. (2023) reviewed FL's medical imaging applications specifically, noting that diagnostic imaging tasks — where data volume and heterogeneity across institutions are both substantial — have become something of a proving ground for the approach. Huang et al. (2019) took a somewhat different angle, showing that clustering patients prior to federated training (rather than treating each hospital as a single undifferentiated client) improved efficiency in predicting mortality and hospital stay duration from distributed records, a refinement worth noting since naive client partitioning is not always the most effective design choice.

Darzidehkalani et al. (2022), writing specifically about radiology, and Hamood Alsamhi et al. (2023), surveying FL's role in pandemic-preparedness contexts more broadly, both arrive at a similar conclusion from different directions: FL is reasonably mature as a training methodology, but its path to routine clinical deployment is still constrained less by the algorithms themselves than by institutional, regulatory, and infrastructural readiness.

4. Privacy-Preserving Mechanisms: Differential Privacy, Secure Aggregation, and Homomorphic Encryption

FL alone reduces but does not eliminate privacy risk — model updates, even without raw data attached, can still leak information about the training set under certain attack conditions. This has motivated a layer of additional protections sitting on top of the basic federated architecture. Differential privacy works by injecting calibrated noise into model updates before they leave a client, bounding how much any individual record can influence the aggregated result (Sav et al., 2022). Secure aggregation, by contrast, addresses a different threat model: it ensures the central server can compute an aggregate of client updates without ever observing any single client's update in isolation, which matters because a curious or compromised aggregator is itself a credible risk in multi-hospital settings. Homomorphic encryption goes further still, allowing computation directly on encrypted values, though typically at a meaningful computational cost (Sav et al., 2022).

Kaissis et al. (2020), in a widely cited piece on medical imaging specifically, argued that these techniques are not interchangeable — they protect against different adversaries and impose different trade-offs, and a framework claiming "privacy-preserving" status needs to

Table 1. Synthesis of Reviewed Studies on Federated Learning, Privacy-Preservation, and Explainability in Healthcare. FL = Federated Learning; DP = Differential Privacy; XAI = Explainable Artificial Intelligence.

be explicit about which threat model it's actually defending against. Butt et al. (2023) illustrated one practical instantiation of this layered approach, combining fog computing with privacy-preserving FL for smart healthcare applications, distributing trust across edge nodes rather than concentrating it at a single aggregator. The broader regulatory backdrop matters here too: Shahid et al. (2022) and Bharati and Podder (2022) both situate these technical mechanisms within the context of HIPAA- and GDPR-style compliance obligations, noting — reasonably — that technical privacy guarantees and legal compliance are related but not identical requirements.

It's worth being honest about the cost side of this equation as well. Gu et al. (2022) examined the trade-offs explicitly, finding that stronger privacy guarantees (tighter differential privacy budgets, in particular) tend to correlate with measurable, if often modest, reductions in model accuracy and fairness across demographic subgroups — a tension that any proposed framework integrating these techniques needs to acknowledge rather than gloss over.

5. Explainable AI for Clinical Trust and Interpretability

If privacy addresses whether data can be shared, explainability addresses something closer to whether the resulting model can be trusted at all — a distinct, and arguably underappreciated, requirement in clinical contexts. Buhrmester et al. (2021) catalogued the broader landscape of explainability methods for black-box models, providing useful grounding for why this matters: deep learning's predictive strength is often inversely related to its transparency, and clinicians are, understandably, reluctant to act on recommendations they cannot interrogate.

SHAP and LIME have emerged as the two most consistently applied techniques in the clinical FL literature reviewed here. Aldughayfiq et al. (2023) applied both methods to interpret deep learning models for retinoblastoma diagnosis, finding that feature-level explanations meaningfully changed how confidently clinicians engaged with model outputs — not simply whether they agreed with the prediction, but whether they felt able to evaluate it on its merits. Minh et al. (2022), in a broader review of XAI methods generally, made a related point: explainability is not a single technique so much as a family of approaches with different assumptions, and the choice of method matters for whether the resulting explanation is actually faithful to the model's internal reasoning, as opposed to merely plausible-looking.

A few studies have pushed past static feature attribution toward more clinically actionable forms of explanation. Düsing and Cimiano (2023) explored counterfactual explanations specifically for sepsis treatment prediction within a federated setting — a notable contribution, since counterfactuals ("the model would have predicted differently if X had been different") arguably map more naturally onto clinical reasoning than SHAP-style feature importance alone. Hoyos et al. (2023) took yet another approach, using fuzzy cognitive maps for dengue-related clinical decision-making within an FL framework, which are interpretable by design rather than requiring a separate post-hoc explanation layer — an architectural choice worth noting as an alternative to bolting XAI onto an otherwise opaque model after the fact. S Band et al. (2023) and Xu et al. (2023), in their respective systematic treatments of interpretability methods in medical AI, both converge on a similar caution: explanations can build misplaced trust just as easily as warranted trust if their limitations aren't communicated clearly to the clinicians relying on them.

6. Blockchain-Enabled Auditability in Federated Systems

Blockchain integration is, of the three technical strands reviewed here, the least mature within healthcare-specific FL literature — though the underlying motivation is straightforward enough. Federated systems involve many independent parties contributing to a shared model over time, and demonstrating after the fact that no party tampered with training logs, model updates, or explanation outputs is a genuinely hard problem without some form of tamper-evident record-keeping. Wu et al. (2023), in a survey dedicated specifically to blockchain-based FL, outlined the general design pattern: a distributed ledger records hashes of model updates (and, in some implementations, associated metadata such as explanation outputs) at each training round, creating an auditable history that no single participant — including the central aggregator — can unilaterally rewrite.

Menegatti et al. (2023) described one applied instance of this combination in healthcare specifically, with the CADUCEO platform supporting federated healthcare facilities through an AI architecture incorporating accountability mechanisms broadly consistent with this approach. It is worth noting, though, that genuinely blockchain-integrated, healthcare-specific FL implementations remain comparatively rare in the literature surveyed here relative to FL-privacy or FL-XAI pairings individually — a gap discussed further in Section 8.

9. Challenges and Limitations Across the Literature

7. Synthesis Matrix of Reviewed Studies

Table 1 summarizes the studies discussed above, organized by their primary technical contribution and clinical domain. Several entries from earlier drafts of this synthesis were removed during preparation because, on verification against the original source, the cited study's actual subject matter did not support the claim being made — a correction noted here in the interest of transparency rather than silently dropped. Removed examples included a citation attributed to FL-based CKD detection that was, on inspection, an unrelated study on database ETL validation; a citation attributed to FL-based liver disease prediction that was actually about non-invasive inflammation biomarkers with no FL component; and a citation attributed to preterm birth prediction that was actually a 6G vehicle-networking paper. Readers and co-authors are encouraged to independently verify the remaining entries against full text before submission, since this check was conducted at the title/abstract level rather than through complete re-reading of every cited source.

8. Toward an Integrated Framework: A Proposed Conceptual Model (PPFBXAIO)

Looking across Table 1, a pattern emerges that's perhaps more interesting than any single study's findings: the literature clusters fairly cleanly into FL-plus-privacy pairings and FL-plus-explainability pairings, with blockchain-based auditability trailing as a comparatively recent and less-integrated addition. Very few of the studies reviewed here attempt all three simultaneously within a single clinical framework, and none in this set combine all three and validate the result across multiple real or realistically simulated hospital sites at once.

That gap is the starting point for PPFBXAIO (Privacy-Preserving Federated Blockchain-logged Explainable AI Optimization), proposed here as a conceptual architecture rather than a tested system. The design, sketched at a high level, would combine: (1) a federated training layer in which each participating hospital trains locally on its own data, following the general pattern established by Brisimi et al. (2018) and extended in domain-specific form by Bebortta et al. (2023); (2) a privacy-preserving layer combining differential privacy and secure aggregation during model update transmission, informed by the threat-model distinctions raised by Kaissis et al. (2020) and the trade-off considerations documented by Gu et al. (2022); (3) an explainability layer generating SHAP-based and, where feasible, counterfactual explanations alongside each prediction, following the approach demonstrated by Aldughayfiq et al. (2023) and Düsing and Cimiano (2023); and (4) a lightweight blockchain logging layer recording hashes of model updates and explanation outputs at each training round, consistent with the general design pattern surveyed by Wu et al. (2023).

It bears repeating that this is a proposal grounded in literature synthesis, not a system that has been built and benchmarked. Its value, such as it is, lies in giving subsequent empirical work a concrete architecture to test against — rather than each new study reinventing some subset of these four components from scratch.

A few limitations recur often enough across the reviewed studies to be worth naming explicitly, rather than treating each as an isolated finding. Communication overhead is probably the most consistently reported practical obstacle — frequent model updates between many hospital nodes and a central aggregator impose real bandwidth and latency costs, and several authors note this becomes a binding constraint at scale rather than a minor inconvenience (Wu et al., 2023). Data heterogeneity across institutions — differing patient populations, equipment, and documentation practices — is a second recurring theme, and one that complicates both model convergence and the interpretability of any single global model's behavior (Sandhu et al., 2023; Darzidehkalani et al., 2022).

The privacy-accuracy-fairness trade-off documented by Gu et al. (2022) deserves particular attention, since it cuts against a tempting but oversimplified narrative in which privacy protections are treated as a "free" addition to an FL pipeline. They are not free, and frameworks proposing to combine multiple privacy mechanisms — as PPFBXAIO does — should be expected to demonstrate, empirically, where on that trade-off curve they land. Finally, the relative scarcity of blockchain-integrated healthcare FL studies noted in Section 6 suggests this remains the least empirically tested of the three technical strands reviewed here, and claims about its benefits in healthcare specifically should be treated as more provisional than the FL-privacy or FL-XAI literature individually.

10. Future Directions

Three directions seem like reasonably natural next steps given the gaps identified above. First, and most directly, empirical validation of integrated frameworks like PPFBXAIO against real or carefully simulated multi-hospital data — ideally with transparent reporting of the privacy-accuracy trade-off rather than accuracy figures reported in isolation. Second, more attention to explanation fidelity rather than just explanation availability — several authors caution that plausible-looking explanations are not the same as faithful ones, and clinical deployment arguably needs the latter (Minh et al., 2022; S Band et al., 2023). Third, longitudinal, real-world deployment studies examining clinician adoption and workflow integration, since nearly everything reviewed here is evaluated in retrospective or simulated settings rather than live clinical use.

Conclusion

This review brings together a fragmented but rapidly growing body of work on federated learning, privacy-preserving computation, and explainable AI in healthcare. Individually, each strand is reasonably mature: federated approaches consistently demonstrate that institutions can collaborate without centralizing patient data, privacy-preserving techniques offer credible, if imperfect, protection against inference risks, and explainability methods meaningfully improve clinician engagement with model outputs. What remains comparatively rare is their integration — few studies attempt all three at once, and fewer still add blockchain-based auditability into the mix. Building on the patterns identified across the reviewed literature, we propose PPFBXAIO as a conceptual framework intended to unify these elements for multi-hospital clinical decision support. The framework is, deliberately, a starting point rather than a finished product; its value will depend on subsequent empirical work testing it against real or realistically simulated multi-institutional clinical data.

Acknowledgements

The authors thank their respective institutions for their support during the preparation of this manuscript. No external funding was received for this review.

Author Contributions

J.A. (Jobayar Alom): Conceptualization, literature search, original draft preparation. M.T.I. (Md Tanzimul Islam): Conceptualization, methodology design, writing — review and editing, corresponding author. M.S.M. (Md Shahriar Masud): Literature search, data curation, writing — review and editing. S.S. (Sudip Saha): Validation, supervision, writing — review and editing. All authors read and approved the final manuscript.

Competing Financial Interests

The authors J.. et al., declare no competing financial interests.

References