Semantic Discovery for Medical Literature Search

Regulatory Framework

Medical literature search sits inside a regulatory perimeter that is rapidly tightening around clinical decision support. The FDA's evolving framework for AI/ML-enabled Software-as-a-Medical-Device (SaMD), articulated in the Predetermined Change Control Plan guidance and the Good Machine Learning Practice principles, requires that clinical AI systems demonstrate transparent reasoning, traceable evidence sourcing, and the ability to support post-market surveillance. The EU AI Act classifies clinical decision support and medical literature synthesis tools that influence diagnosis or treatment under Annex III §5 (access to and enjoyment of essential services, including healthcare), imposing high-risk obligations: risk management, data governance, technical documentation, logging, transparency, human oversight, and accuracy/robustness/cybersecurity requirements that attach to every output the system produces.

Methodological standards layer onto the regulatory base. ICH Good Clinical Practice (GCP) governs the conduct of clinical trials whose results populate the literature; CONSORT defines the reporting standard for randomized trials; PRISMA defines the reporting standard for systematic reviews and meta-analyses; GRADE defines the framework for rating evidence certainty across outcomes. Cochrane methodology operationalizes risk-of-bias assessment for systematic reviews. Each of these standards encodes the principle that evidence carries differential weight based on design, conduct, and reporting quality, and clinical decisions are expected to reflect that differential.

The corpora themselves carry implicit governance. PubMed/MEDLINE indexes the peer-reviewed literature with MeSH terms encoding subject and study type. The Cochrane Library curates systematic reviews to defined methodological standards. ClinicalTrials.gov records prospective trial registrations whose existence constrains post-hoc reporting bias. FDA drug approval documents and EMA EPARs contain regulator-reviewed evidence summaries. A search system that flattens these sources into a single ranked list discards the governance the corpora encode.

Architectural Requirement

The regulatory and methodological framework imposes architectural requirements that retrieval systems built for general web search do not satisfy. First, the system must distinguish evidence grades structurally, not as a post-hoc filter. A systematic review of randomized trials is not interchangeable with an observational cohort or a case report, and the discovery process must reflect the hierarchy in its traversal logic, not in a faceted filter the user must remember to apply.

Second, the system must maintain persistent clinical context. Clinical questions are rarely resolved in a single search; they unfold over hours or days as the clinician integrates findings, formulates new sub-questions, and reconciles conflicting evidence. Stateless retrieval forces the clinician to reconstruct context with every query, which discards the accumulated reasoning that defines clinical thought.

Third, the system must produce traversal lineage. EU AI Act Article 12 logging requirements, FDA SaMD post-market surveillance expectations, and the documentation requirements of evidence-based practice each require that the path from clinical question to evidence-supported answer be reconstructible. Lineage is not an audit add-on; it is a first-class output of the discovery process.

Fourth, the system must support cross-corpus traversal without flattening. A single clinical question may need synthesized evidence from PubMed, Cochrane, ClinicalTrials.gov, FDA labels, and EMA EPARs, and the synthesis must preserve the source-specific governance each corpus carries. Fifth, the system must support human oversight in the form the EU AI Act requires: the clinician must be able to inspect the reasoning, contest individual traversal steps, and override conclusions, and the system must record the override in the lineage.

Why Procedural Compliance Fails

Current medical search systems satisfy procedural expectations without delivering substantive evidence governance. PubMed's relevance ranking blends term-frequency, recency, and citation signals into a list that does not encode the GRADE hierarchy. MeSH publication-type filters allow restriction to randomized trials or systematic reviews, but the restriction operates as exclusion rather than as weighted traversal: a clinician who filters to RCTs loses access to observational evidence on adverse effects, mechanistic insights from basic science, and the regulatory context that FDA and EMA dossiers provide. The filter trades one form of bias for another.

Embedding-based semantic search, increasingly deployed as a layer over PubMed and proprietary medical corpora, intensifies the problem. Embedding similarity rewards textual proximity and stylistic overlap, neither of which correlates with evidentiary weight. A well-written case report whose abstract uses the clinician's query terms in close proximity will rank above a relevant systematic review that uses the same concepts in different language. The embedding model is doing what it was trained to do; the failure is architectural, not algorithmic.

Retrieval-augmented generation pipelines compound the failure. An LLM generating a clinical summary from retrieved passages cannot recover evidence grade from the passages themselves because the passages have been stripped of corpus context, study-design metadata, and risk-of-bias signals during retrieval. The output reads fluently and may cite sources, but the citations are weighted by retrieval similarity, not by GRADE certainty, and the clinician reading the output cannot distinguish a citation backed by a Cochrane review from one backed by a single underpowered observational study.

Clinical guideline development tools and systematic-review-assistance platforms have begun to address evidence-grade differentiation, but they treat it as a downstream classification rather than a traversal primitive. Studies are retrieved and then graded; the grading does not steer the retrieval. A search that retrieves the wrong studies cannot be rescued by post-hoc grading, and the clinician's time budget is consumed by manual triage that the system should have performed structurally.

Statelessness is the failure mode that wastes the most clinical time. A physician researching a complex case across consultations, specialist input, and treatment iteration is forced to re-search the same ground repeatedly because the search system does not remember what was already established. The accumulated evidence assessment exists only in the clinician's notes, where it is not inspectable, not transferable to colleagues, and not auditable for the documentation evidence-based practice requires.

What AQ Primitive Provides

The Adaptive Query semantic-discovery primitive replaces ranked retrieval with governed traversal. The clinical question is instantiated as a persistent discovery object carrying the patient context, the question being asked, the corpora authorized for the question, the evidence-grade weighting profile appropriate to the question type, and the accumulated assessment to date. Traversal is trust-scoped: each step through the corpus respects the GRADE hierarchy and the source-specific governance encoded in the corpus, and each step contributes to the lineage that the discovery object carries.

The trust scope is not a filter; it is a weighting that steers traversal. A treatment-efficacy question weights systematic reviews and randomized trials most heavily but does not exclude observational evidence when randomized evidence is unavailable, mechanistic evidence when the clinical question requires understanding of action, or regulatory evidence when labeling and post-market surveillance are decision-relevant. The weighting profile is question-specific because the appropriate evidence mix is question-specific: adverse-effect questions require observational and post-market data that randomized trials underpower; mechanistic questions require basic science that the clinical hierarchy nominally subordinates but that clinical reasoning depends on.

The persistent discovery object accumulates context across sessions. A clinician returning to a complex case finds the discovery object in the state it was left, with the questions already addressed, the questions still open, the evidence already assessed, and the contradictions still unresolved. The object is shareable: a specialist consulted on the case receives the discovery object rather than a search summary, and contributes to the same lineage rather than initiating a parallel one. The object is auditable: the documentation requirements of evidence-based practice are satisfied by the lineage the object carries, not by retrospective reconstruction.

Cross-corpus traversal preserves source governance. The discovery object can traverse from a Cochrane review to the underlying trials registered on ClinicalTrials.gov, then to the FDA approval dossier that incorporated those trials, then to the EMA EPAR that adjudicated the same evidence, with each transition recorded and each source's governance preserved in the lineage. The synthesis is governed rather than flattened.

Human oversight is structurally supported. The clinician can inspect any traversal step, contest the trust weighting applied to a specific source, override the system's evidence assessment, and the override is recorded in the lineage with the clinician's reasoning. The EU AI Act's human-oversight requirement is satisfied by the architecture rather than bolted on as a UI feature.

Compliance Mapping

The semantic-discovery artifact maps onto the regulatory perimeter as a structural fit. Against FDA AI/ML SaMD expectations, the discovery object's lineage provides the transparent reasoning and traceable evidence sourcing the Good Machine Learning Practice principles require, and the object's persistent state supports the post-market surveillance the Predetermined Change Control Plan framework anticipates. Against EU AI Act Annex III §5 high-risk obligations, the lineage satisfies Article 12 logging, the trust-scoped traversal satisfies Article 10 data governance, the discovery object's inspectability satisfies Article 13 transparency, and the human-oversight architecture satisfies Article 14.

Against ICH GCP, CONSORT, PRISMA, and GRADE, the trust-scoped traversal operationalizes the methodological hierarchy these standards encode, replacing the manual application of methodology with structural enforcement. Against the corpus-specific governance of PubMed, Cochrane, ClinicalTrials.gov, FDA labels, and EMA EPARs, the cross-corpus traversal preserves source governance through the lineage rather than discarding it through flattening.

Adoption Pathway

Adoption proceeds along a clinically-grounded path. Stage one is the individual clinician deploying semantic discovery as a research workspace for complex cases, where the persistent discovery object replaces the scattered notes and re-searched ground that current practice produces. Stage two is the clinical service line adopting shared discovery objects for case conferences, tumor boards, and morbidity and mortality reviews, where the lineage produces the documentation these settings require.

Stage three is the hospital system integrating semantic discovery with the electronic health record, binding discovery objects to patient encounters and producing the evidence trail that accreditation and quality programs increasingly demand. Stage four is the clinical guideline development organization adopting semantic discovery as the methodological substrate for guideline production, where the trust-scoped traversal and lineage replace the labor-intensive manual workflows that currently dominate guideline methodology.

Stage five is regulatory recognition: as semantic discovery accumulates use in regulated clinical decision support, FDA and EMA review pathways can recognize the architecture as satisfying the AI/ML SaMD and AI Act high-risk requirements structurally, reducing the documentation burden on each individual deployment. The path is incremental, clinically-grounded, and operates within the existing methodological and regulatory framework rather than requiring its replacement.