acceptodds
Under review as a conference paper at ICLR 2027

FCTG: Fuzzy Cognitive Topology Graph for Uncertainty- Aware Medical Diagnosis

Abstract

Knowledge-graph-augmented large language models (KG-LLMs) are promising for medical diagnosis, but most medical KGs still collapse symptoms and symptom–disease relations into binary links or single-point probabilities, losing the linguistic vagueness and statistical uncertainty inherent in clinical reasoning. We propose the Fuzzy Cognitive Topology Graph (FCTG), a KG-RAG framework that represents both patient-side symptom memberships and graph-side symptom–disease relations as trapezoidal fuzzy numbers, composes them through fuzzy arithmetic, and organizes evidence into confidence nested (-cut) subgraphs. Patient-side fuzzy numbers encode the clarity and severity of symptom descriptions, whereas graph-side fuzzy numbers encode the statistically estimated strength of symptom–disease evidence. FCTG then uses fuzzy-number multiplication and aggregation to rank candidate diseases and assemble structured context for final LLM diagnosis without modifying the LLM itself. On the DDXPlus triage dataset, FCTG achieves 88.0%/98.0% Top-1/Top-3 accuracy, exceeding the strongest scalar KG baseline for each metric by 22.0 and 10.0 percentage points, respectively. On the Gretelai free-text diagnosis dataset, FCTG reaches 74.0% Top-1 accuracy, ahead of the strongest baseline by 4 points. Extensive experiments validate the effectiveness of FCTG in medical diagnosis scenarios.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Fuzzy Knowledge Graphs, Retrieval-Augmented Generation, Fuzzy Membership, Semantic Symptom Membership, Large Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_3P8vOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Fuzzy Cognitive Topology Graph (FCTG), a knowledge-graph (KG) augmented LLM for diagnoses that proposes to represent symptoms and symptom-disease relationships as fuzzy numbers that can be composed under fuzzy arithmetic. The graph topology naturally captures uncertainty and normalized positive likelihood ratios and combines these quantities through approximate fuzzy multiplication and addition. Disesases are subsequently ranked by the resulting centroids, and a restricted candidate set is supplied to a frozen diagnostic LLM to complete the full KG-LLM setup.

Critical strengths

On the DDXPlus cases the model has SOTA results

The use of the fuzzy math + KG is highly creative, has valuable compositional features, and I find the construction to be very well motivated to the problem space itself of relating symptoms to diagnoses

Critical weaknesses

The benchmarks themselves are artificial (literally synthetic) and bear only a passing relationship to real medical use cases and distributions.

The baselines are sampled from similarly architected systems: MedRAG, MedGraphRAG, etc... while leaving out baseline LLM comparisons (how does Sonnet 5.5 or GLM5.3 perform on this?)

Questions for authors

How is scalar patient confidence computed and what is the scoring rule for the “scalar edge or product” (trapezoidal width) configuration?

Why not have more standard baselines for the diagnostic reasoning field (MedR-Bench, SD Bench, HealthBench Pro, etc...) ?

Reviewer_GhWXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces the a knowledge graph RAG framework (FCTG) to handle clinical uncertainty in LLM-based medical diagnosis. Instead of using binary links or single-point probabilities, FCTG represents both graph-level symptom-disease associations and patient-level symptom descriptions as trapezoidal fuzzy numbers. At inference time, an alpha-cut threshold filters out low-confidence symptoms and graph edges, and the remaining fuzzy weights are multiplied, summed per disease, and defuzzified via their centroid to rank candidate diagnoses for final LLM reranking. Experiments shows the accuracy gain from this method.

Critical strengths
  1. Separating statistical reliability in KG from linguistic vagueness in patient complaints addresses a genuine failure mode of standard KG-RAG pipeline, where incidental or weakly phrased symptoms are weighted equally to definitive ones. As shown in Table 4, scalar baselines degrade by 11% to 22% when moving from mild to confusable symptom descriptions on DDXPlus, whereas FCTG drops by only 6%.

  2. The evaluation of the paper is pretty comprehensive. It evaluates LLM backbones Qwen3.8-Flash and Qwen2.5-72B-Instruct and provide candid analyses of where their method does and does not help. Notably, they show that alpha-cut pruning does not improve end-to-end runtime because LLM latency dominates graph scoring, and they clearly explain why gains on Gretelai are modest due to sparse training edges and already-deterministic patient complaints.

Critical weaknesses
  1. Appendix F reveals that FCTG's diagnostic prompt includes an explicit "FCTG-specific uncertainty guidance" instruction telling the LLM not to let mild, vague, or incidental findings overturn the KG prior, whereas this instruction is omitted from the G1–G3 scalar baselines. Because the ablation study in Table 5 removes the entire LLM reranker rather than ablating this prompt guidance independently, it is impossible to tell how much of the DDXPlus gain comes from fuzzy scoring versus the extra prompt instructions.

  2. On Gretelai in Appendix A, FCTG uses a 330-edge graph that retains singleton edges, whereas G1–G3 are restricted to a 270-edge graph that drops those 60 singleton edges. Since missing graph edges cause the majority of Gretelai failures (as in Table 12), giving FCTG 60 additional edges directly confounds its lead over G1.

  3. While the paper emphasizes end-to-end trapezoidal fuzzy arithmetic as its core methodological advance over scalar Fuzzy Cognitive Maps, Table 5 shows that replacing trapezoidal widths with scalar centroids reduces Top-1 accuracy only slightly on DDXPlus and Gretelai.

  4. Evaluation is limited to only 100 test cases per split across two synthetic benchmarks from a single fixed run, without variance estimates or statistical significance indicator.

Questions for authors

I'd love to see how my concerns in weaknesses are addressed.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.