Ontology-based knowledge representation for bone disease diagnosis: a foundation for safe and sustainable medical artificial intelligence systems

Loan Dao; Ngoc Quoc Ly

REVIEW 4 major objections 5 minor 2 references

Ontology-based knowledge representation for bone disease diagnosis: a foundation for safe and sustainable medical artificial intelligence systems

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that a hand-built bone-disease ontology with SWRL rules can guide segmentation-classification, visual question answering, and multimodal deep-learning models, keeping AI diagnoses interpretable and reliable.

desk verdict A well-structured design proposal for ontology-guided bone disease AI whose formal rules and equations fail to parse and whose validation claims are internally inconsistent. read the letter →

arxiv 2506.04756 v1 pith:NT3WRFS4 submitted 2025-06-05 cs.AI cs.CVeess.IV

Loan Dao , Ngoc Quoc Ly This is my paper

classification cs.AIcs.CVeess.IV

keywords bonediseaseontologyknowledgerepresentationmedicalartificialintelligenceSWRLrulesvisualquestionansweringmultimodaldeeplearningclinicalinterpretabilityontology-guided

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper is trying to establish that a hand-crafted ontology of bone diseases, expressed in Description Logic (a formal language for classes and relationships) and SWRL rules (if-then clinical logic), can serve as the knowledge backbone for three kinds of medical AI: VLM-guided segmentation-classification, visual question answering, and multimodal deep learning. The authors argue that when the ontology constrains the model's attention, features, and output paths, the resulting diagnoses are more reliable and clinically interpretable than black-box approaches. The central claim is architectural and theoretical: digitized medical knowledge, mapped to standard terminology and organized into hierarchical classes and relationships, provides a reproducible foundation for safe and sustainable medical AI. If the framework is right, bone-disease diagnosis systems could be built with transparent reasoning paths and extended to other medical domains. The paper states that experimental validation remains pending because of dataset and computational constraints.

What carries the argument

The central object is the BoneDx Ontology, a formal knowledge structure $O = (C, R, A)$ in Description Logic (ALC): a set of concepts $C$, relations $R$, and axioms $A$. It encodes disease hierarchies, e.g. $D_i \sqsubseteq D_j \sqcap M$ for automatic malignancy inheritance, and multimodal feature integration $F \sqcap \exists hasFeature.(I \sqcup L \sqcup T)$. The reasoning machinery is a set of SWRL rules, e.g. $BoneDisease(?d) \wedge hasSymptom(?d, Pain) \wedge hasLocation(?d, FemoralNeck) \wedge hasAgeGroup(Patient, Elderly) \rightarrow suspectDiagnosis(?d, FemoralNeckFracture)$, which convert the ontology into executable diagnostic logic. Ontology-guided attention, $Attention(Q,K,V)=softmax(QK^T/\sqrt{d_k})V$, is used to align learned feature weights with domain knowledge. Together these components carry the paper's claim that knowledge constraints, not just data statistics, drive diagnosis.

What would settle it

Run the BoneDx Ontology with its SWRL rules on a labeled collection of bone-disease cases and check agreement: if applying the rules to cases with known diagnoses produces wrong conclusions (for example, the femoral-neck fracture rule firing for a young patient with hip pain from another cause), the claim that ontology-guided constraints improve diagnostic reliability fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that medical knowledge can be formalized into a reusable ontology, the BoneDx Ontology, with seven primary classes (Disease, AnatomicalStructure, Diagnosis, Symptom, Treatment, Patient, Cause), object and data properties such as hasLocation, hasDiagnosis, hasSymptom, and SWRL rules that automate clinical reasoning, such as inferring a suspected femoral-neck fracture from pain at that location in an elderly patient. This ontology is inserted into three diagnostic pipelines: a hierarchical segmentation-classification network that uses VLM prompts, an ontology-enhanced VQA system that returns structured clinical answers, and a multimodal deep learner fusing imaging, laboratory, and text data under ontological constraints. The claimed result is that these constraints improve AI decision reliability while preserving clinical interpretability, because every output follows a structured reasoning path grounded in standardized medical terminology.

Load-bearing premise

The paper assumes that the hand-built ontology and its SWRL rules, derived from standard medical knowledge and input from a single hospital, are complete and correct enough to constrain AI diagnosis.

Editorial extensions

If this is right

Ontology-guided constraints should improve the reliability of bone-disease AI decisions by enforcing anatomical and clinical relationships during segmentation, classification, and reasoning.
The same standardized structure can extend beyond bone diseases, because the modular design and terminology mapping are reusable in other medical domains.
VQA responses can be made explainable by routing reasoning through the ontology, producing answers that cite anatomical location, fracture pattern, and patient-specific risk factors.
The 1,247-case hospital dataset, dominated by trauma cases (70.2%) and lower-limb injuries (52.8%), provides a preliminary basis for future validation.
Clinical deployment would require system certification, healthcare staff training, and integration with existing hospital infrastructure, as the paper identifies.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If validated, the SWRL rules can be read as individual testable hypotheses: each rule claims a specific symptom-location-age combination implies a diagnosis, so auditing the AI becomes auditing a finite set of clinical rules.
The architecture implies a division of labor opposite to current end-to-end trends: clinical relationships are fixed a priori by the ontology, and the neural networks learn only perception and alignment, making the system's ceiling the quality of the hand-built rules.
A natural next experiment is an ablation that compares ontology-guided models with unguided baselines on the same hospital data, and measures how often the ontology overrides or corrects the model's raw prediction.
The ontology's property axiom $hasLocation \circ isPartOf \sqsubseteq hasLocation$ suggests a formal way to test spatial generalization: if a model locates a lesion in a bone that is part of a larger structure, the transitivity rule should propagate that localization consistently across anatomical levels.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

I read the ontology-based bone disease diagnosis paper (arXiv:2506.04756). The first thing to know: the paper openly admits that experimental validation is pending—this is stated in the abstract and in Section 2.2.3. That honesty is to its credit. What the paper does well is assemble a plausible architecture: an ontology-guided hierarchical network with VLM prompts, an ontology-enhanced VQA module, and a multimodal fusion model. The literature review is competent, and the citation pattern (including self-citations for prior segmentation/classification work) is not abusive.

The problems are in the formal core. The stress-test note is accurate: Equation (1) 'D_i ⊑ D_j ⊓ M' implies every disease subclass is malignant if M represents malignancy, which cannot be intended. The SWRL rules in Table 4 are not valid SWRL; the osteoarthritis rule is written in Manchester OWL syntax, and the fracture rule uses 'hasAgeGroup(Patient, Elderly)' where Patient is a class, not an individual, and the property is undefined. Table 5's individual axiom is not a Description Logic axiom. These are not cosmetic typos—they are load-bearing because the whole framework's reliability claim depends on executable rules. The paper also contradicts itself: Section 3 says 'validation using 1,247 cases ... demonstrates practical applicability,' while the abstract and Section 2.2.3 say validation remains pending. That internal inconsistency needs explicit resolution.

The novel piece here is the domain-specific integration, not the conceptual framework—combining ontologies with ML and VLMs is well established. As submitted, the central claims about reliability and interpretability are unsupported and the formal machinery is unsound as specified. A Protégé/HermiT parsing and consistency check would quickly expose the issues.

Who is this for? A reader interested in medical AI design proposals might skim it, but anyone citing it would need to wait for a corrected and validated version. I would not cite it. I would send it to peer review only because the topic is relevant and a rigorous reviewer could force the authors to fix the formal errors and run a real consistency test—but the likely outcome is reject-and-resubmit, not acceptance.

In short: this is a blueprint, not a result. It could become useful after substantial revision and actual validation.

Referee Report

4 major / 5 minor

Summary. The paper proposes an ontology-based framework for bone disease diagnosis, combining a bone disease ontology (BoneDx) with SWRL rules, a hierarchical neural network with VLM-based segmentation-classification, an ontology-enhanced VQA system, and a multimodal deep learning model. It claims three theoretical contributions toward safe and sustainable medical AI and states that experimental validation remains pending. The manuscript includes descriptive dataset statistics from 1,247 hospital cases, tables of ontology classes and properties, and a set of rules and axioms that are supposed to encode clinical reasoning, but it contains no experiments, ablations, or quantitative evaluation.

Significance. If the framework were sound and validated, it could contribute to a genuinely useful direction: making medical AI interpretable and knowledge-guided. The collaboration with a clinical hospital and the concrete dataset context are positive elements. The proposed modular architecture (ontology-guided attention, VQA, multimodal fusion) is a plausible program of research. However, as submitted, the paper does not deliver what it claims. There are no experimental results despite Section 3 asserting demonstrated practical applicability, and the formal artifacts that constitute the core theoretical contribution contain load-bearing logical and syntactic errors. The manuscript does not ship machine-checked proofs, reproducible code, or parameter-free derivations; Equations (3) and (4) are standard formulas, while Equation (1) and the rule tables are not valid as written. The central reliability and interpretability claims are therefore unsupported.

major comments (4)

[2.2.1, Eq. (1)] Equation (1), D_i ⊑ D_j ⊓ M, with M representing malignancy status (benign/malignant), implies that every specific disease category is necessarily a subtype of malignancy. This is clinically false for benign diseases and is not a harmless notation shortcut: the paper explicitly states in Section 2.2.1 that this formalization 'ensures automatic malignancy determination upon disease classification.' The formal basis of the framework is therefore internally inconsistent as written.
[2.3.2, SWRL fracture rule and Table 4] The fracture diagnosis rule is not a valid SWRL rule. It contains hasAgeGroup(Patient, Elderly), where Patient is a class, not an individual, and hasAgeGroup is not defined in Table 3 (the property table). The Osteoarthritis Diagnosis Rule in Table 4 is written in Manchester OWL syntax (e.g., 'BoneDisease and (hasSymptom some Pain) and (hasSymptom some LimitedMobility) and (hasAge some xsd:integer[>= 50])'), not in SWRL, and it mixes class expressions with a data range. These rules cannot be parsed by standard SWRL engines or reasoners such as Protégé/HermiT, so the purported executable reasoning layer does not exist as specified.
[2.3.2, Table 5] The Individual Axiom example, 'Fracture(x) ∧ hasLocation(x,Femur) → requiresImaging(x,XRay)', is neither a valid Description Logic individual axiom nor a valid SWRL rule: it uses untyped constants and conjunction in the antecedent, and '∧' is not SWRL syntax. This matters because Table 5 is presented as the reasoning foundation for automated diagnostic support. A simple Protégé/HermiT parse-and-consistency check would expose these issues, and the paper as it stands cannot be implemented or evaluated.
[Section 3 and Table 6] Section 3 states that 'validation using 1,247 cases from Ho Chi Minh City Hospital demonstrates practical applicability,' but Section 2.2.3 explicitly says 'experimental validation remains as future work,' and Section 2.3.3 describes the dataset as a preliminary foundation. The only reported numbers are descriptive distributions (trauma 70.2%, lower limb 52.8%). No diagnostic accuracy, sensitivity, specificity, ablation, or comparison with a baseline is provided. The assertion in Table 6 that 'Ontology-guided constraints improve AI decision reliability' is an empirical claim with no supporting experiment, and it is contradicted by the paper's own statement that validation remains pending.

minor comments (5)

[2.3.2] The cross-reference 'Error! Not a valid bookmark self-reference.' appears after the SWRL fracture rule and should be replaced with the actual table number.
[Tables 2 and 3] Notation is inconsistent: 'BoneDisease' appears with different capitalizations and spacing (e.g., 'bone diseases', 'BoneDisease'), and Table 3 uses 'AnatomicalStr ucture' with a space. Consider unifying terminology and class names.
[2.3.1, Fig. 1] The figure is referenced but not described in enough detail in the text for readers to follow the three-layer architecture; a full caption or a textual breakdown of the layers would help.
[2.3.3, Algorithm 1] Algorithm 1 is a trivial four-step preprocessing list (remove special characters, lowercase, tokenize, map to standard terminology) and does not substantiate the label 'Medical Text Normalization' as a methodological contribution; either expand it with concrete mapping details or remove it.
[References] The paper cites Horridge and Bechhofer (2011) and Shearer et al. (2008) for OWL API and HermiT, respectively, but it does not report using either tool. If the ontology was not actually parsed or reasoned over with these tools, the citations are misleading.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is a theoretical ontology framework with explicitly pending validation, and no claimed prediction is equivalent to its inputs by construction.

full rationale

The paper does not walk a derivation chain from fitted inputs to predicted outputs. Its three contributions are architectural proposals, and the text explicitly disclaims empirical validation: 'experimental validation remains pending due to current dataset and computational resource limitations' (Abstract) and 'the framework's experimental validation remains as future work' (Section 2.2.3). The ontology classes and SWRL rules are the inputs to the proposed system, not quantities fitted from data and then renamed as predictions. Equation (1)'s 'automatic malignancy determination' is a standard reasoning consequence of the declared subclass axioms (e.g., Osteosarcoma ⊑ MalignantTumor), not a circularly predicted result. The Section 3 statement that 'validation using 1,247 cases ... demonstrates practical applicability' is unsupported because no evaluation metrics or experiments are reported, but that is a correctness/evidence gap, not a definitional reduction. The self-citations (Dao & Ly 2023, 2024) are background literature and are not load-bearing for the ontology's correctness. The 'ontology-guided constraints improve AI decision reliability' benefit stated in Table 6 is an unverified assertion; the absence of validation is not circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 2 invented entities

The central claim rests on the correctness and completeness of the hand-crafted ontology and SWRL rules, which are introduced ad hoc for this paper. There are no fitted parameters because there are no experiments. The paper's axioms are domain assumptions about the adequacy of DL and SWRL for representing clinical reasoning, and the invented entities (BoneDx Ontology, ontology-guided attention) lack independent evidence in this submission.

free parameters (2)

Ontology class hierarchy and SWRL rules
The ontology and its rules are hand-crafted by the authors based on one hospital's clinical input; no empirical fitting or validation is done in the paper. These are effectively free design parameters that determine the system's behavior.
Thresholds and data split (not specified)
No experimental parameters are reported because no experiments were run. The paper leaves all such parameters unspecified.

assumptions (3)

domain assumption Description Logic (ALC) is an adequate formal foundation for medical diagnosis knowledge.
The paper assumes that ALC-based ontologies with SWRL rules can capture the clinical reasoning needed for bone disease diagnosis. This is stated in Section 2.2.1 but not justified with evidence.
domain assumption The ontology's class hierarchy and SWRL rules are clinically correct and complete.
The entire framework depends on the accuracy of the hand-built ontology (Tables 2, 3, 4, 5). No clinical validation or review is provided.
domain assumption VLM-based segmentation and VQA can be effectively tuned with ontology-guided prompts.
The framework assumes that VLMs can follow ontology-derived prompts reliably enough for medical decision support, but the paper only provides a conceptual equation (Equation 3) and no experiments.

invented entities (2)

BoneDx Ontology
purpose: A structured representation of bone disease knowledge for guiding AI diagnosis.
The ontology is introduced in this paper, and its correctness is asserted without external evidence or validation. It is not tied to any external standard benchmark in the paper.
Ontology-guided attention mechanism
purpose: An attention mechanism claimed to align learned features with domain knowledge.
This is a conceptual adaptation of standard attention (Equation 4) with no implementation or evaluation; it is a proposal, not an established entity.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Ontology-based knowledge representation for bone disease diagnosis: a foundation for safe and sustainable medical artificial intelligence systems." pith.science (2026). https://pith.science/paper/NT3WRFS4

@misc{pith2026250604756,
  author       = {Pith},
  title        = {Pith review of: Ontology-based knowledge representation for bone disease diagnosis: a foundation for safe and sustainable medical artificial intelligence systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NT3WRFS4}},
  note         = {Machine review of arXiv:2506.04756}
}

read the original abstract

Medical artificial intelligence (AI) systems frequently lack systematic domain expertise integration, potentially compromising diagnostic reliability. This study presents an ontology-based framework for bone disease diagnosis, developed in collaboration with Ho Chi Minh City Hospital for Traumatology and Orthopedics. The framework introduces three theoretical contributions: (1) a hierarchical neural network architecture guided by bone disease ontology for segmentation-classification tasks, incorporating Visual Language Models (VLMs) through prompts, (2) an ontology-enhanced Visual Question Answering (VQA) system for clinical reasoning, and (3) a multimodal deep learning model that integrates imaging, clinical, and laboratory data through ontological relationships. The methodology maintains clinical interpretability through systematic knowledge digitization, standardized medical terminology mapping, and modular architecture design. The framework demonstrates potential for extension beyond bone diseases through its standardized structure and reusable components. While theoretical foundations are established, experimental validation remains pending due to current dataset and computational resource limitations. Future work will focus on expanding the clinical dataset and conducting comprehensive system validation.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

2 extracted references · 1 canonical work pages

[1]

Chen, H., Wang, Y., & Li, K. (2024). Vision -language models in medical imaging: A systematic review. Medical Image Analysis, 89, 102952. https://doi.org/10.1016/j.media.2024.102952 Dao, L., & Ly, N. Q. (2023). A comprehensive study on medical image segmentation using deep neural networks. International Journal of Advanced Computer Science and Application...

arXiv 2024
[664]

https://www.mdpi.com/2075-4418/13/4/664 Mou, Y., Zhang, X., & Wang, H. (2024). Knowledge Graph -enhanced Vision -to-Language Multimodal Models for Radiology Report Genera tion. In Proceedings of the Extended Semantic Web Conference (ESWC 2024). https://2024.eswc-conferences.org/wp-content/uploads/2024/05/77770446.pdf Noy, N. F., & McGuinness, D. L. (2001)...

work page arXiv 2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Chen, H., Wang, Y., & Li, K. (2024). Vision -language models in medical imaging: A systematic review. Medical Image Analysis, 89, 102952. https://doi.org/10.1016/j.media.2024.102952 Dao, L., & Ly, N. Q. (2023). A comprehensive study on medical image segmentation using deep neural networks. International Journal of Advanced Computer Science and Application...

arXiv 2024

[2] [664]

https://www.mdpi.com/2075-4418/13/4/664 Mou, Y., Zhang, X., & Wang, H. (2024). Knowledge Graph -enhanced Vision -to-Language Multimodal Models for Radiology Report Genera tion. In Proceedings of the Extended Semantic Web Conference (ESWC 2024). https://2024.eswc-conferences.org/wp-content/uploads/2024/05/77770446.pdf Noy, N. F., & McGuinness, D. L. (2001)...

work page arXiv 2024