How can proof-grounded patient-specific clinical explanations be generated from knowledge graphs?

Proof-grounded means an inspectable rule certificate

In this paper, "proof" does not mean an external theorem prover has certified logical soundness or clinical truth. It means each exported rule application records four audit-relevant fields: the premises consumed by the rule, a rule identifier, the resulting conclusion, and a human-readable reason or provenance reference. Trace completeness is defined operationally as the presence of those required fields in the exported artifact.

That definition matters because it sets the verification target. A deterministic report is generated from trace-level conclusions rather than directly from raw workbook cells, so a sentence or action can be audited backwards to a conclusion, from the conclusion to a rule application, and from the rule application to patient and knowledge facts. The engine also requires every emitted recommendation to resolve to a policy identifier and every cited knowledge feature to resolve to packaged YAML provenance.

Availability, representation, and activation are orthogonal states

The engine's central design choice is to keep three things apart that are often collapsed into one missingness label. Measurement availability is a binary fact about whether a measurement exists. Representation coverage describes how completely an observed feature is encoded. Disease-specific activation, the information-gate value IG_D(p, f), is a separate quantity in the range [0,1] and is explicitly not an availability flag.

The paper states the consequence directly: an IG value of zero does not imply the feature is unobserved. An observed measurement can contribute zero disease-specific activation, and an observed feature can simultaneously be partially represented. The corpus contained 86,783 observed feature instances with IG = 0 and 139,949 observed instances with partial representation, which the authors present as the stress cases showing why a single generic "missing or weak evidence" label would discard materially different computational states.

Traces are built from exported evidence, not recomputed

For each patient-disease pair, the explanation engine constructs the trace from already exported feature-level evidence rather than reconstructing the upstream diagnostic calculation. A patient value is first associated with its mapped feature and disease-specific role. Availability is then read independently from representation coverage and activation. The ordering is deliberate: an IG value is interpreted only after the system knows whether the underlying measurement was observed, and coverage is treated as a property of representation rather than acquisition.

The upstream system is frozen. It exports patient values, feature mappings, clinical roles, disease-level scores, evidence coverage, and the disease-specific information-gate value, and the explanation layer treats these as immutable facts. Disease-level scores are therefore evidence to be explained, not quantities the engine recomputes. The paper notes that a change to the explanation renderer cannot feed back into the diagnostic score.

Only unresolved evidence plus a policy entry licenses a test

The recommendation boundary is deliberately narrower than the descriptive trace. The trace may describe observed support, observed non-activation, partial representation, or unresolved evidence, but only the final category can participate in a confirmatory-test rule, and then only when a disease/feature/test mapping exists in the packaged policy. Low coverage, partial representation, or a language model cannot independently create a test recommendation.

This separates describing uncertainty from authorizing an action. The paper gives the example that an observed feature with IG = 0 can be described as observed but non-activating, and cannot be relabelled "missing"; partial representation can be surfaced as uncertainty in encoded coverage but cannot itself authorize a test. The design is described as intentionally conservative: a clinically sensible action absent from the policy is not silently invented but requires an explicit knowledge or policy update outside the patient-level explanation run.

Corpus-scale checks passed, with one localized exception

Across five usable workbooks covering 6,720 patients, 20,160 patient-disease traces, and 1,021,440 feature-evidence rows, IG-range validity and knowledge provenance resolution were both 100% (1,021,440/1,021,440), numerical cross-sheet fidelity was 100% (120,960/120,960), and exported logical trace completeness and deterministic-report trace completeness were both 100% (20,160/20,160).

Availability-representation consistency was 99.7028% (1,018,404/1,021,440). The paper stresses that the 3,036 disagreements were not heterogeneous failures but were confined to three systematic feature-cohort combinations: 1,018 VOM_manifestation records in Dengue traces and 1,018 VOM_progression records in Malaria traces from the same 1,018-patient workbook, plus 1,000 MAL_manifestation records in Dengue traces from a 1,000-patient workbook. In every case the feature-evidence export marked the measurement as observed with IG = 0, while the confidence audit recorded zero representational coverage and status = missing.

The authors characterize this as an availability-representation disagreement between exported structures, not evidence that the engine converts the measurement to unobserved. Because the engine reads availability from the feature-evidence Observed field, a zero-coverage confidence-audit record does not by itself satisfy the unobserved premise of the critical-incomplete rule. The paper states this is a code-path conclusion about the current engine, not evidence that the legacy export is harmless for every external consumer, since software interpreting status = missing as acquisition missingness could reach a different state. The disagreement is retained as a source-data and schema defect rather than normalized away.

The seeded audit exercised the executable pathway

A separate seeded 25-patient end-to-end audit completed without execution failure and passed all pre-specified proof-completeness, recommendation-licensing, source-resolution, and availability/representation checks. Mean execution time varied across the five workbooks from approximately 14.8 to 52.0 seconds per audited patient in the recorded experimental environment; the paper states these measurements characterize the audited implementation rather than a general deployment benchmark.

For an illustrative patient (ID 152), the frozen upstream Dengue trace reported evidence probability 0.1179, Hallmark/Major cosine similarity 0.7062, relative magnitude 0.6737, disease score 0.4993, and evidence coverage 0.3715. Hemoconcentration and mild thrombocytopenia were observed high-role evidence with IG = 1.000 and approximately 0.953 respectively, while leukopenia/neutropenia remained observed despite IG = 0. Other high-role evidence was partially represented, and NS1/PCR evidence was genuinely unobserved, making it eligible for confirmatory-test licensing only through policy.

The paper is explicit that the two audits answer different questions. Corpus-scale checks provide breadth over more than one million feature-evidence rows and are stronger for detecting systematic export inconsistencies; the seeded engine audit verifies that the current executable pathway can construct complete trace and report artifacts for sampled patients and is stronger for detecting execution, policy-resolution, or report-construction failures. Neither substitutes for clinical validation.

Structural validity is not semantic or clinical validity

The paper separates three validity layers. Structural validity asks whether exported values, provenance, trace fields, and policy identifiers are internally consistent, and this is the principal target of the present experiments. Semantic validity asks whether the encoded medical roles, thresholds, and rules faithfully represent the intended clinical knowledge. Clinical validity asks whether using the resulting explanation improves decisions or patient outcomes. Passing the first layer does not establish the latter two, and the authors note that a perfectly traceable system can still expose an incorrect encoded rule.

The study is cross-sectional and does not evaluate prognosis, patient outcomes, clinician decision quality, or prospective clinical utility. The 25-patient engine audit is smaller than the corpus-level exported-structure audit and is not presented as exhaustive end-to-end clinical validation. The optional language realization layer is intentionally excluded from the reported pass rates; the paper makes no claim that the constrained architecture outperforms unrestricted LLM reporting, calling that a separate empirical question. The zero-byte influenza workbook is excluded before analysis and is not included in any denominator.

The audit value sits in the zero-activation and partial cases

The most decision-relevant reading of this paper is not the headline 100% figures, which describe structural checks on exported artifacts, but the two large populations those checks were designed to protect: 86,783 observed instances with IG = 0 and 139,949 observed instances with partial representation. These are exactly the cases where a conventional renderer could collapse distinct computational states into one "missing/weak evidence" label, and where the engine's explicit predicates keep observed non-activation, partial encoding, and genuine non-observation distinguishable.

The second priority is the recommendation boundary. Because a test action requires both a genuinely unobserved policy-listed feature and a matching policy entry, the auditable question for any deployment is not whether the trace is complete but whether the packaged policy contains the action a clinician would consider appropriate. The paper's own framing supports this ordering: traceability makes an incorrect encoded rule inspectable rather than clinically correct, so the next step it identifies is clinician review of rule and policy content and prospective evaluation, not further structural auditing.

Sources used in this answer