PaperlayerView exact revision on arXiv

Source-linked AI summary

Training-Free Clinical Reasoning through Medical Ontologies and Cognitive Mapping: A Symbolic-Probabilistic Knowledge Graph Framework

Surajit Das

arXiv:2609.35298v1cs.AI

TL;DR

Clinical prediction commonly learns patient-variable-outcome associations, while this paper investigates reasoning from explicit medical knowledge without fitting diagnostic mappings to cohort labels. CKG Reasoner uses candidate-specific evidence representations and separates ranking from label-free cohort assignment; across six retrospective cohorts, positive-class F1 ranged from 0.634 to 0.996. The findings describe knowledge-grounded evidence separation, not prospective clinical validity or comparative superiority.

  • Problem

    Clinical findings have different support, contradiction, importance, specificity, and decisive roles, while learned feature–outcome mappings need not explicitly preserve these distinctions.

    Source: §1 · p. 1

  • Method

    CKG Reasoner maps patient observations to explicit candidate-specific medical knowledge and separates evidence-based ranking from K-means assignment using derived evidence coordinates.

    Source: §7 · p. 17

  • Results

    Across six retrospective cohorts, positive-class F1 scores were 0.996, 0.634, 0.936, 0.917, 0.695, and 0.842, with full partition-decision coverage in all six.

    Source: §7 · p. 17

  • Takeaways & Limitations

    The results characterize knowledge-grounded evidence separation and its dependence on cohort evidence regimes, rather than prospective clinical performance or superiority over outcome-fitted classifiers.

    Source: §5 · p. 15, §7 · p. 17

    The framework is cross-sectional, and matched classifier comparisons, clinical utility, prospective validation, and clinical validation of FOL/LLM explanations remain outstanding.

    Source: §6 · p. 17

Abstract

from arXiv · show

Most clinical prediction systems learn patient-variable-outcome associations; we investigate a training-free diagnostic paradigm mapping patient observations to explicit medical knowledge. CKG Reasoner integrates candidate-specific Evidence Feature Nodes, patient-reference matching, a bounded Information Gate, knowledge-weighted evidence accumulation, disease similarity, and decisive clinical rules. Missing-aware normalization and coverage auditing distinguish absent from unavailable evidence. Candidate ranking is separate from outcome-label-independent K-means clustering, which uses four derived evidence coordinates (evidence strength, relative magnitude, directional similarity, and evidence completeness), not raw predictors or targets, to derive cohort-level assignments. Across six retrospective cohorts - four dengue (N = 1000, 1523, 989, 1018), malaria (N = 2190), and influenza (N = 4569) - a uniform, label-free, cohort-fitted K = 2 protocol yielded positive-class F1 scores of 0.996, 0.634, 0.936, 0.917, 0.695, and 0.842, and all-record accuracies of 0.996, 0.558, 0.914, 0.893, 0.707, and 0.906, respectively, with full partition-decision coverage using the frozen package and disease-specific knowledge representations. Neither scoring nor clustering uses outcome labels. Logistic regression provides a supervised baseline. Influenza incorporates confirmatory molecular PCR and is not independent pre-test prediction. Results characterize knowledge-grounded evidence separation, auditability, and sensitivity, not prospective clinical validity or comparative superiority. FOL/LLM-based clinical explanation remains unevaluated.

1. Introduction

Clinical diagnosis must integrate findings with different evidential roles, while learned classifiers need not preserve distinctions such as contradiction, decisive evidence, and unresolved information. CKG Reasoner tests whether explicit medical knowledge can support fixed, auditable reasoning without fitting the diagnostic mapping to cohort outcome labels.

1. Introduction

  • Clinical findings can provide weak or strong support, contradiction, decisive confirmation, or exclusion, and differ in importance and disease specificity.Outcome-fitted models may perform well in a population but may degrade under unmodeled distribution shift or confounding; knowledge-based systems can be brittle with incomplete or ambiguous evidence.Source: §1 · p. 1
  • “Training-free” means outcome-label-independent: patient labels do not train, fit, fine-tune, calibrate, or optimize the reasoner.The package and disease-specific knowledge representations were developed independently of the six evaluation cohorts and frozen before evaluation; clinical parameters encode specified medical knowledge on fixed ordinal scales.Source: §1 · p. 1
  • The framework separates candidate ranking, decisive clinical rules, evidence-completeness auditing, and cohort-level unsupervised assignment as distinct operations.Ranking is relative support, not a calibrated posterior probability or prospectively validated diagnosis; missing evidence is not treated as contradiction.Source: §1 · p. 2
  • The study asks whether explicit medical knowledge can be operationalized as a fixed reasoning process without fitting diagnostic mappings to cohort labels, rather than whether it outperforms supervised classifiers.The authors identify coherence, auditability, perturbation robustness, evidence-removal sensitivity, and provenance as primary questions.Source: §1 · p. 2
  • Across six separate disease-specific cohorts, uniform label-free K = 2 clustering evaluated frozen disease-specific representations, not a three-class differential-diagnosis benchmark.The study explicitly does not claim predictive superiority over outcome-fitted classifiers.Source: §1 · pp. 2–3
  • The architecture integrates candidate-conditioned symbolic–probabilistic reasoning, evidence provenance, and a separate unsupervised diagnostic-assignment pathway.The paper distinguishes implemented capabilities from future extensions including interactive evidence acquisition, real-time support, uncertainty-region validation, and FOL/LLM explanation validation.Source: §1 · pp. 2–3

2. Related Work and Positioning

Prior work establishes probabilistic rules, ontology-based diagnosis, knowledge graphs, and learned patient-specific systems. The paper positions its contribution as integrating distinct evidence semantics within a fixed, auditable, patient-activated reasoning architecture.

2.1. From expert rules to probabilistic and ontology-based diagnosis

  • Probability–rule hybrids and ontology-based systems already combine probabilistic inference, fuzzy rules, semantic similarity, or symptom dependencies.The paper’s distinction is graph-structured separation of evidence semantics and decisive clinical roles, not probability-plus-rules or semantic uncertainty alone.Source: §2.1 · p. 3

2.2. Knowledge graphs for diagnostic reasoning

  • Knowledge-graph diagnostic systems include weighted paths, embedding-refined paths, and exclusion-aware graph scoring, so transparent exclusion-aware scoring is not claimed as unique.The present framework distinguishes multiple evidence quantities rather than primarily representing uncertainty through weighted relations.Source: §2.2 · p. 3

2.3. EHR-centered, patient-specific, and learned clinical graphs

  • Recent patient-specific systems use learned, graph-neural, agentic, or language-model components, whereas this work activates a fixed knowledge-driven procedure without outcome-label parameter estimation.The paper assumes a prespecified clinical knowledge model rather than addressing fragmented-data integration as an EHR-oriented graph workflow.Source: §2.3 · p. 3

2.4. Neuro-symbolic and LLM-guided reasoning

  • FOL serialization and LLM-based verbalization are prospective extensions; the study evaluates numerical evidence and rule traces, not those generated explanations.Neuro-symbolic and structured-knowledge LLM systems are described as retaining learned or generative components.Source: §2.4 · p. 3

2.5. Closest methodological gap

  • The methodological gap is preserving multiple evidence meanings as separate operational quantities within one outcome-label-independent, patient-activated architecture, rather than claiming novelty for individual components.These quantities include importance, specificity, patient–reference agreement, support, contradiction, decisive roles, and an inference trace tied to the diagnostic result.Source: §2.5 · p. 4

3.1. Clinical Knowledge Representation

The reference graph encodes disease-specific clinical knowledge while keeping it distinct from how a patient’s observations match that knowledge. Candidate-conditioned Evidence Feature Nodes separate diagnostic meaning and evidential strength from patient–reference correspondence.

3.1.1. Reference knowledge graph

  • The reference graph links observed clinical features and contextual variables to candidate diseases as a cross-sectional evidence structure, not a temporal progression model.Its sources include guidelines, pathology, biomarkers, imaging, epidemiological evidence, and expert knowledge.Source: §3.1.1 · p. 4
  • The representation separates patient–reference correspondence from disease-knowledge attributes such as role, importance, support, and contradiction.A feature’s patient match is distinct from its encoded disease specificity and meaning.Source: §3.1.1 · p. 4
  • Diagnostic role encodes candidate-specific disease specificity and direction, whereas clinical importance reflects practical diagnostic importance and can vary independently.A specific feature may be absent or poorly matched, while a strongly matched feature may be nonspecific; importance can depend on availability, reliability, disease stage, and context.Source: §3.1.1 · p. 4
  • Support strength represents positive evidential strength for an appropriately matched present feature, independently of its specificity or clinical importance.Features with the same role and importance may differ in support because of consistency, context dependence, or alternative explanations.Source: §3.1.1 · pp. 4–5
  • Contradiction magnitude represents negative evidence independently of positive support, allowing agreement and disagreement to have asymmetric consequences.A characteristic uncommon finding may strongly support disease when present but contradict little when absent; an expected but nonspecific finding can have the reverse pattern.Source: §3.1.1 · p. 5
  • An Evidence Feature Node stores a candidate-specific feature and its canonical state, while patient observations activate against that state during inference.The same observation can be interpreted against different canonical states and diagnostic attributes across candidate disease graphs.Source: §3.1.1 · pp. 5–6

3.2. Hybrid Diagnostic Reasoning

CKG Reasoner applies frozen, outcome-label-independent medical knowledge to candidate-specific patient–reference mappings rather than fitting diagnostic associations to cohort outcomes. Its operational scoring separates feature correspondence, evidence accumulation, disease similarity, ranking, and evidence completeness, with missing evidence handled separately from observed absence.

3.2.1. Training-free design

  • The ontology, disease-specific representations, and inference parameters were frozen before evaluation; outcome labels were reserved for retrospective evaluation, and K-means used four derived coordinates rather than raw predictors or labels.Centroids were fitted independently within each cohort, making the retrospective assignment procedure transductive rather than an independent out-of-sample prediction test.Source: §3.2.1 · p. 6

3.2.2. Ontology-grounded cognitive mapping

  • Candidate-conditioned cognitive mapping gives each observation disease-specific meaning by matching it to that candidate’s canonical states, roles, and evidence semantics.The decomposition is intended to make errors localizable, but training-free reasoning does not guarantee clinical correctness because it depends on the supplied knowledge and mapping functions.Source: §3.2.2 · p. 6

3.2.3. Operational scoring

  • Component activations are bounded knowledge-engineering compatibility scores, not probabilities or measures of prevalence, specificity, importance, support, or contradiction.The same patient value may produce different activations under different candidate diseases because their canonical reference states differ.Source: §3.2.3 · p. 7
  • Feature-level scoring combines directional agreement and relative magnitude in a bounded Information Gate; missing or unavailable evidence is not represented as a zero gate value.An Information Gate value of 1 denotes perfect patient–canonical correspondence for that feature.Source: §3.2.3 · p. 7
  • Candidate evidence is normalized against a canonical self-reference over the same observed feature support, so unavailable features neither count as negative evidence nor enlarge the denominator.The resulting normalized evidence score is not calibrated as a probability and can be negative or exceed one.Source: §3.2.3 · p. 8
  • Diagnostic Evidence Coverage measures observed, cohort-evaluable diagnostic evidence separately from candidate scoring, and its Confidence Score can affect clustering but not candidate ranking.The four-coordinate representation retains evidence strength, disease-level magnitude and direction, and completeness; unresolved clinically important features are also reported in the audit trace.Source: §3.2.3 · pp. 8–9

3.2.4. Hard clinical rules as implemented

The framework applies candidate-specific decisive rules after scoring, audits evidence completeness separately from ranking, and builds inference traces from activated knowledge-graph features. Formal and language-model explanation pathways are proposed extensions; the executed experiments do not evaluate them.

3.2.4. Hard clinical rules as implemented

  • Pathognomonic and exclusionary triggers require candidate-specific patient–reference agreement, but no exclusionary EFNs were encoded in the supplied knowledge files for the six evaluated runs.Both trigger thresholds were set to 0.8; the exclusionary override was available in code but was not empirically triggered.Source: §3.2.4 · p. 9

3.2.5. Candidate ranking, decisive rules, and evidence-completeness audit

  • Candidate ranking uses an additive relative score with explicit hard-rule overrides, while Diagnostic Evidence Coverage is audited separately and can influence K-means through the fourth representation coordinate.Missing evidence is distinguished from observed absence and is not converted into contradictory evidence; unresolved Pathognomonic, Hallmark, Major, and Exclusionary EFNs are exported.Source: §3.2.5 · p. 9

3.2.6. Patient-specific graph extraction and inference-faithful explanation

  • Patient-specific graph extraction activates disease-specific Evidence Feature Nodes, and each candidate trace exposes the evidence, rules, scores, and provenance that produce support.The same observation may enter multiple candidate graphs with different canonical representations and evidence values.Source: §3.2.6 · p. 10

3.2.7. Formal explanation and reasoning-to-language extension

  • The proposed explanation object organizes evidence facts, numerical contributions, rule states, completeness, decisions, and provenance, while keeping numerical operations outside pure first-order logic.The proposed object distinguishes candidate ranking from partition-derived diagnostic assignment.Source: §3.2.7 · p. 10

3.2.8. Prospective constrained clinical-report generation

  • The experiments do not evaluate an LLM verbalizer, end-to-end claim grounding, clinician-rated explanations, or clinical report quality.Structured, provenance-linked inputs and deterministic verification are described as preparatory components for a possible future verbalizer.Source: §3.2.8 · p. 10, §3.2.7 · p. 10

3.2.9. End-to-end diagnostic workflow

  • The executed workflow produces knowledge-grounded inference traces and K-means partition assignments; conversion to formal explanations and verified LLM reports remains an unevaluated extension.Cohort A remains a dengue-versus-non-dengue task; negative records are not relabeled as other diseases.Source: §3.2.9 · p. 10, §3.2.7 · p. 10

3.3. Cross-Sectional One-Layer Formulation

The cross-sectional formulation aggregates patient evidence without inferring temporal progression, then separates disease ranking from cohort-level partition assignment. The evidence score is not a time-normalized index.

3.3. Cross-Sectional One-Layer Formulation

  • With no diagnostic-time attributes, the patient score is Best_Score, a diagnostic evidence score rather than a time-normalized DEI.Dengue records are cross-sectional, with one record per patient.Source: §3.3 · p. 10
  • Figure 1 distinguishes executed inference and partition assignment from available explanation interfaces and a proposed, unevaluated FOL/LLM clinical-report workflow.Dashed arrows indicate extension paths, not experimental results.Source: Figure 1 · p. 11
  • Patient observations activate candidate-specific evidence and accumulate into candidate rankings, while evidence-completeness auditing remains a distinct operation.The workflow then applies K-means to a post-reasoning representation for cohort-level diagnostic tags.Source: Figure 2 · p. 11

3.4. Training-Free Multi-Dataset Evaluation

The evaluation applies a frozen, outcome-label-independent reasoner to six clinical cohorts using deterministic, disease-specific mappings. A uniform K = 2 protocol evaluates post-reasoning assignments, with evidence completeness reported separately from candidate ranking.

3.4.1. Evaluation principle

  • Outcome labels are excluded from package construction, schema mapping, candidate scoring, ranking, and K-means fitting, entering only after assignment for retrospective evaluation.K-means uses standardized four-coordinate vectors, not the scalar ranking score.Source: §3.4.1 · p. 11

3.4.2. Evaluation datasets

  • The evaluation covers six cohorts across dengue, malaria, and influenza, with two additional dengue cohorts using the frozen package and existing Dengue representation.The first four reported cohorts are listed as Bangladesh dengue, hematology dengue, malaria, and Thailand influenza.Source: §3.4.2 · p. 11

3.4.3. Disease-specific feature mapping

  • Dataset variables are deterministically mapped to disease-specific EFNs, preserving stage identity, normalizing supported values, and distinguishing missing from observed-negative findings.The dengue mapping audit and additional dengue mappings document matched and skipped variables; outcome fields are not used as evidence.Source: §3.4.3 · pp. 11–12, Table 3 · p. 12

3.4.4. Primary endpoints and partition protocol

  • All cohorts use K = 2 on standardized four-coordinate post-reasoning representations, with positive-class F1 as the main F1 measure and full partition-decision coverage.No intermediate partition is generated under the two-cluster protocol.Source: §3.4.4 · p. 12

3.4.5. Evidence-completeness reporting

  • Outputs report Evidence Coverage, CS, confidence information, unresolved critical-evidence fields, and feature-level traces without changing evidence accumulation or candidate ranking.CS is also the fourth standardized clustering coordinate and can affect partition assignments.Source: §3.4.5 · p. 12

3.5. Comparators and Controlled Experiments

The study provides contextual comparator information and controlled robustness analyses, while distinguishing these from matched predictive comparisons. Several proposed evaluation targets remain untested.

3.5.1. Comparator scope

  • The supplied supervised logistic-regression results are contextual rather than a matched comparison because their held-out test subsets differ from CKG's retrospective partition evaluation.The results do not establish predictive superiority for either approach.Source: §3.5.1 · p. 12
  • The comparator audit considers executable implementations, artifacts, input requirements, and compatibility with the cross-sectional cohorts; published implementation availability does not establish direct applicability.Reproduction of published results and adaptation to these datasets are distinct tasks.Source: §3.5.1 · p. 12, Table 4 · p. 13

3.5.2. Ablation and robustness protocol

  • Secondary experiments test component ablations, evidence withholding, gate sensitivity, and initialization stability without fitting ontology weights or activation parameters to outcome labels.These analyses cover only compatible K = 2 configurations; quantitative evidence-acquisition and blinded explanation-quality experiments are not reported.Source: §3.5.2 · p. 13

4. Results

Under a uniform K = 2 protocol, the six cohorts show varied retrospective partition metrics with full coverage, alongside evidence of robustness and cohort-dependent sensitivity. These results characterize knowledge-grounded assignments, not prospective clinical validity or matched predictive superiority.

4. Results

  • The reported logistic-regression F1 values use different populations and decision procedures from the CKG metrics, so they are not a direct head-to-head test.Prospective clinical performance, calibration, and clinical utility were not evaluated.Source: §4 · p. 13, §4.1 · p. 14, Table 5 · p. 14

4.2. Bootstrap uncertainty and seed stability

  • For B, C, D, and F, bootstrap intervals quantify uncertainty, and available partitions were unchanged across 20 K-means seeds (ARI=1.000), not independent validation.Comparable uniform-protocol bootstrap intervals were not reported for A and E.Source: §4.2 · p. 14

4.3. Component ablation

  • Component-ablation effects varied by cohort: removing Pevidence changed Dataset F more substantially, while Dataset C and D partitions remained stable under several single-coordinate removals.Dataset D's Pevidence-only separation should be interpreted in light of confirmatory laboratory evidence.Source: §4.3 · p. 14, Table 7 · p. 15

4.4. Evidence-gate, partition, and missing-evidence sensitivity

  • Gate-setting sensitivity was limited in tested cohorts, while 50% evidence withholding reduced performance in several cohorts and exposed the limits of F1 alone under class imbalance.The common K = 2 protocol does not establish that two clusters are clinically optimal.Source: §4.4 · pp. 14–15

4.5. Influenza confirmatory-evidence ablation

  • Removing PCR alone preserved Dataset D's partition, but removing PCR and rapid-antigen evidence reduced F1 from 0.842 to 0.555, supporting interpretation as post-test integration rather than pre-test prediction.The combined removal is a pre-test-like evidence-withholding stress test, not a prospectively designed pre-test study.Source: §4.5 · p. 15

4.6. Clustering diagnostics and ranking separation

  • K-means partitions standardized evidence coordinates rather than the scalar disease-ranking score, keeping cohort assignment operationally distinct from candidate ranking.Assignments are ordered without reference outcomes during fitting.Source: §4.6 · p. 15

5. Discussion

The discussion frames CKG Reasoner as an auditable, label-independent alternative for knowledge-grounded reasoning, not as a demonstration of predictive superiority. Retrospective performance varies across evidence regimes, with tested stability and evidence sensitivity alongside translational capabilities that remain unvalidated.

5.1. Training-free ontology-grounded reasoning

  • CKG separates patient-specific evidence semantics and provenance from outcome-fitted parameter estimation, making individual contributions inspectable.Candidate ranking and cohort-level assignment are distinct; reported metrics evaluate unsupervised partitions of four post-reasoning coordinates.Source: §5.1 · p. 15

5.2. Cross-cohort variability and baseline context

  • Positive-class F1 ranged from 0.634 to 0.996 across six cohorts, and the reported baselines do not establish matched comparative superiority.The authors relate variation to heterogeneous evidence and evaluation conditions, without claiming these fully explain it.Source: §5.2 · pp. 15–16
  • Dataset B relied mainly on routine hematological findings with limited disease specificity; its CKG F1 was 0.634 versus logistic-regression F1 0.682 under separate protocols.The comparison is descriptive and does not establish the cause of the performance limitation.Source: §5.2 · pp. 15–16
  • Dataset E achieved F1 0.936 and Dataset F 0.917 under additional dengue schemas, without outcome-based knowledge modification; these results do not establish prospective or external-site validity.Across cohorts, performance differed with evidence regimes; influenza's F1 of 0.842 reflects post-test integration because PCR is encoded as pathognomonic evidence.Source: §5.2 · p. 16

5.3. Robustness and evidence dependence

  • Partitions for cohorts B, C, D, and F were unchanged across 20 K-means seeds, while evidence withholding and component ablations showed cohort-dependent sensitivity.At 50% withholding, F1 declined to 0.447 for malaria and 0.498 for influenza; removing both PCR and rapid-antigen evidence reduced influenza F1 from 0.842 to 0.555.Source: §5.3 · p. 16
  • Evidence-gate sensitivity was predominantly stable within tested parameter ranges, but stability under population shift, alternative knowledge representations, or prospective deployment was not established.Comparable secondary results for cohorts A and E were unavailable where not reported.Source: §5.3 · p. 16

5.4. Explainability, uncertainty, and translational extensibility

  • The inference trace preserves distinct clinical quantities and provides a foundation for investigating explanations, uncertainty regions, interactive evidence acquisition, and patient-wise reassessment.The six-cohort evaluation does not validate these capabilities, real-time deployment, or FOL/LLM-generated clinical explanations.Source: §5.4 · p. 16
  • Diagnostic Evidence Coverage describes observed cohort-evaluable evidence, while the Confidence Score contributes to clustering but not candidate ranking.The Confidence Score is not a calibrated probability of diagnostic correctness.Source: §5.1 · p. 15, §5.4 · p. 16

6. Limitations

The framework's scope is constrained by encoded knowledge and feature mappings, limited modeling of feature interactions, and fixed representations that do not adapt automatically to clinical shifts. Its cohort-level partitions and evidence coverage do not establish calibrated uncertainty or prospective clinical utility.

Clinical knowledge and evidence representation

  • Reasoning depends on accurate, complete medical knowledge and consistent feature mappings; population-specific ranges, normalization, and activation boundaries need further study.Although exclusionary reasoning is implemented, it was not empirically evaluated in the six cohorts.Source: §6 · p. 17

Hierarchical and cooperative evidence reasoning

  • The framework does not explicitly model hierarchical, conditional, or cooperative feature interactions, whose combined significance may differ from individual evidence contributions.The authors identify resolving conflicting observations and incorporating clinically justified dependencies as future work.Source: §6 · p. 17

Knowledge updating and distribution shift

  • Fixed knowledge representations do not automatically adapt to changes in populations, laboratory practices, or disease manifestations.Knowledge updating would require provenance, clinical verification, and version control to preserve auditability.Source: §6 · p. 17

Explicit uncertainty and information acquisition

  • Cohort-level partitioning does not establish a clinically validated uncertainty region or distinguish insufficient, conflicting, and uncertain evidence assignments.The authors call for abstention and clinically relevant information acquisition rather than decisions based on cluster geometry alone.Source: §6 · p. 17

Clinical translation and explanation

  • Diagnostic Evidence Coverage measures evidence availability, not calibrated certainty; its Confidence Score can influence clustering despite being excluded from candidate ranking.Matched supervised comparisons, calibration, clinical utility, prospective selective-decision evaluation, temporal reasoning, and clinical validation of FOL/LLM explanations remain outstanding.Source: §6 · p. 17

7. Conclusion

CKG Reasoner separates knowledge-grounded candidate reasoning from cohort-level diagnostic assignment, and evaluates the approach across six retrospective cohorts. The results do not establish prospective clinical validity or predictive superiority over outcome-fitted classifiers.

7. Conclusion

  • The architecture separates evidence-grounded reasoning and provenance from cohort-level K-means assignment, without fitting the scorer to outcome labels.K-means uses a four-coordinate post-reasoning representation rather than directly classifying from candidate-ranking scores.Source: §7 · p. 17
  • Positive-class F1 scores were 0.996, 0.634, 0.936, 0.917, 0.695, and 0.842 across six cohorts, with full partition-decision coverage in all six.The cohorts comprised four dengue datasets, malaria, and influenza; influenza included confirmatory testing and is not independent pre-test prediction.Source: §7 · p. 17
  • The retrospective results establish neither prospective clinical validity nor predictive superiority over outcome-fitted classifiers.The authors identify independent and temporal validation, matched clinical-task comparisons, and clinician evaluation of the FOL/LLM explanation extension as future priorities.Source: §7 · p. 17
Loading 2609.35298v1…