An ontology for Age-Related Macular Degeneration using ophthalmologists and language models Adrian Groza1 , Anca Marginean1 and Simona Delia Nicoara2 1 Department of Computer Science, Technical University of Cluj-Napoca, 400114 Cluj-Napoca, Romania 2 Department of Ophthalmology, β€œIuliu Hatieganu” University of Medicine and Pharmacy, 400012 Cluj-Napoca, Romania Abstract We aim to support monitoring of the current guidelines and scientific evidence in the management of Age-Related Macular Degeneration (AMD) in order to augment retinal specialists to develop a clinically oriented and consensual protocol for therapeutic approaches for AMD. First, we are engineering an ontology for AMD retinal condition using information from literature, related medical ontologies and domain knowledge from ophthalmologists. Second, we augment the knowledge engineer capabilities to populate and enrich the ontology using structured knowledge extracted from medical literature with the GPT-3 language model. Third, we perform reasoning to signal to the ophthalmologist differences or inconsistencies among different clinical studies, protocols or therapeutic approaches. Keywords medical ontologies, age-related macular degeneration, conflict detection, reasoning First, as an example of axioms from the AMD ontology, consider the classifications of AMD in Table 1, where πΏπ‘Žπ‘Ÿπ‘”π‘’π·π‘Ÿπ‘’π‘ π‘’π‘› ≑ π·π‘Ÿπ‘’π‘ π‘’π‘› βŠ“ βˆƒβ„Žπ‘Žπ‘ π‘†π‘–π‘§π‘’. β‰₯ 125πœ‡π‘š, π‘€π‘’π‘‘π‘–π‘’π‘šπ·π‘Ÿπ‘’π‘ π‘’π‘› ≑ π·π‘Ÿπ‘’π‘ π‘’π‘› βŠ“ βˆƒβ„Žπ‘Žπ‘ π‘†π‘–π‘§π‘’. β‰₯ 63πœ‡π‘š βŠ“ βˆƒβ„Žπ‘Žπ‘ π‘†π‘–π‘§π‘’. β‰₯ 63πœ‡π‘š, respectively π‘†π‘šπ‘Žπ‘™π‘™π·π‘Ÿπ‘’π‘ π‘’π‘› ≑ π·π‘Ÿπ‘’π‘ π‘’π‘› βŠ“ βˆƒβ„Žπ‘Žπ‘ π‘†π‘–π‘§π‘’. ≀ 63πœ‡π‘š. One issue is that these axioms do not always correspond to clinical practice. For instance an eye with a drusen measured by an AI algorithm at 124ΞΌm (i.e. slightly below the 125ΞΌm limit) is classified according to the definition as a π‘€π‘’π‘‘π‘–π‘’π‘šπ·π‘Ÿπ‘’π‘ π‘’π‘›, and hence an πΈπ‘Žπ‘Ÿπ‘™π‘¦π΄π‘€π·, but the ophthalmologist still treats the disease as an 𝐼 π‘›π‘‘π‘’π‘Ÿπ‘šπ‘’π‘‘π‘–π‘Žπ‘‘π‘’π΄π‘€π·. To map the clinical practice we are also considering axioms in Fuzzy Description Logic. The AMD ontology reuses con- cepts and relations from BioVerbNet (https://github.com/cambridgeltl/bioverbnet) and medical ontologies, e.g.: (i) 𝐴𝑀𝐷 βŠ‘ 𝐼 π‘›β„Žπ‘’π‘Ÿπ‘–π‘‘π‘’π‘‘π‘…π‘’π‘‘π‘–π‘›π‘Žπ‘™π·π‘¦π‘ π‘‘π‘Ÿπ‘œπ‘β„Žπ‘¦ βŠ‘ π‘€π‘’π‘›π‘‘π‘’π‘™π‘–π‘Žπ‘›π·π‘–π‘ π‘’π‘Žπ‘ π‘’ βŠ‘ 𝐻 π‘’π‘šπ‘Žπ‘›π·π‘–π‘ π‘’π‘Žπ‘ π‘’; (ii) 𝐼 π‘›β„Žπ‘’π‘Ÿπ‘–π‘‘π‘’π‘‘π‘…π‘’π‘‘π‘–π‘›π‘Žπ‘™π·π‘¦π‘ π‘‘π‘Ÿπ‘œπ‘β„Žπ‘¦ βŠ‘ 𝐼 π‘›β„Žπ‘’π‘Ÿπ‘–π‘‘π‘’π‘‘π‘‰ π‘–π‘‘π‘Ÿπ‘’π‘œπ‘’π‘ π‘…π‘’π‘‘π‘–π‘›π‘Žπ‘™π·π‘–π‘ π‘’π‘Žπ‘ π‘’ βŠ‘ π‘…π‘’π‘‘π‘–π‘›π‘Žπ‘™π·π‘–π‘ π‘œπ‘Ÿπ‘‘π‘’π‘Ÿ βŠ‘ πΈπ‘¦π‘’π·π‘–π‘ π‘œπ‘Ÿπ‘‘π‘’π‘Ÿ. Second, we enrich the AMD ontology with structured data automatically extracted from scientific studies [1] and clinical trials. Recent advances in Natural Language Understanding can complement the studies conducted by humans on reviewing literature and recent scientific evidence (e.g. [2]). Consider querying a learned model like GPT-3 (https://bit.ly/3e3icZQ) in Table 2). The used prompt was: ”A table summarizing the associations of morphological features with disease activity”. One the one hand we were fascinated on the easiness to obtain such structured data. On the other hand, in line with G. Marcus (https://cacm.acm.org/blogs/ SWAT4HCLS 2023: The 14th International Conference on Semantic Web Applications and Tools for Health Care and Life Sciences Envelope-Open adrian.groza@cs.utcluj.ro (A. Groza) GLOBE http://users.utcluj.ro/~agroza/ (A. Groza) Orcid 0000-0003-0143-5631 (A. Groza) Β© 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). CEUR Workshop Proceedings http://ceur-ws.org ISSN 1613-0073 CEUR Workshop Proceedings (CEUR-WS.org) Table 1 Sample of definitions and classifications scales for AMD Epidemiological classification (Wisconsin grading) πΈπ‘Žπ‘Ÿπ‘™π‘¦π΄π‘€π· π‘Š ≑ 𝐴𝑀𝐷 βŠ“ βˆƒβ„Žπ‘Žπ‘ π΅π‘–π‘œπ‘šπ‘Žπ‘Ÿπ‘˜π‘’π‘Ÿ.(πΏπ‘Žπ‘Ÿπ‘”π‘’π·π‘Ÿπ‘’π‘ π‘’π‘› βŠ” π‘…π‘’π‘‘π‘–π‘›π‘Žπ‘™π‘ƒπ‘ π‘’π‘’π‘‘π‘œπ‘‘π‘Ÿπ‘’π‘ π‘’π‘›βŠ” π‘ƒπ‘–π‘”π‘šπ‘’π‘›π‘‘π‘Žπ‘Ÿπ‘¦π΄π‘π‘›) πΏπ‘Žπ‘‘π‘’π΄π‘€π· π‘Š ≑ 𝑁 π‘’π‘œπ‘£π‘Žπ‘ π‘π‘’π‘™π‘Žπ‘Ÿπ΄π‘€π· βŠ” πΊπ‘’π‘œπ‘”π‘Ÿπ‘Žπ‘β„Žπ‘–π‘π΄π‘‘π‘Ÿπ‘œπ‘π‘¦ Basic clinical classification 𝑁 π‘œπ΄π‘”π‘’π‘–π‘›π‘”πΆβ„Žπ‘Žπ‘›π‘”π‘’π‘  𝐢 ≑ βˆ€β„Žπ‘Žπ‘ π·π‘Ÿπ‘’π‘ π‘’π‘›.βŠ₯ βŠ“ βˆ€β„Žπ‘Žπ‘ π΄π‘π‘›.Β¬π‘ƒπ‘–π‘”π‘šπ‘’π‘›π‘‘π‘Žπ‘Ÿπ‘¦π΄π‘π‘› 𝑁 π‘œπ‘Ÿπ‘šπ‘Žπ‘™π΄π‘”π‘’π‘–π‘›π‘”πΆβ„Žπ‘Žπ‘›π‘”π‘’π‘  𝐢 ≑ βˆ€β„Žπ‘Žπ‘ π·π‘Ÿπ‘’π‘ π‘’π‘›.π‘†π‘šπ‘Žπ‘™π‘™π·π‘Ÿπ‘’π‘ π‘’π‘› βŠ“ βˆ€β„Žπ‘Žπ‘ π΄π‘π‘›.Β¬π‘ƒπ‘–π‘”π‘šπ‘’π‘›π‘‘π‘Žπ‘Ÿπ‘¦π΄π‘ AREDS simplified severity scale points π‘†π‘’π‘£π‘’π‘Ÿπ‘–π‘‘π‘¦0 ≑ βˆ€β„Žπ‘Žπ‘ π΅π‘–π‘œπ‘šπ‘Žπ‘Ÿπ‘˜π‘’π‘Ÿ.Β¬πΏπ‘Žπ‘Ÿπ‘”π‘’π·π‘Ÿπ‘’π‘ π‘’π‘› βŠ” βˆ€π‘β„Žπ‘Žπ‘›π‘”π‘’π‘ .Β¬π‘ƒπ‘–π‘”π‘šπ‘’π‘›π‘‘ π‘†π‘’π‘£π‘’π‘Ÿπ‘–π‘‘π‘¦1 ≑ βˆƒβ„Žπ‘Žπ‘ π΅π‘–π‘œπ‘šπ‘Žπ‘Ÿπ‘˜π‘’π‘Ÿ.Β¬πΏπ‘Žπ‘Ÿπ‘”π‘’π·π‘Ÿπ‘’π‘ π‘’π‘› βŠ” (= 1)π‘β„Žπ‘Žπ‘›π‘”π‘’π‘ .π‘ƒπ‘–π‘”π‘šπ‘’π‘›π‘‘ Table 2 Extracting structured information on morphological features using language models (i.e. GPT3) Review Feature Association with disease activity Mowatt et al. (2014) OCT unlikely to be cost-effective for diagnosis/monitoring Schmid-Erfurth et al. (2016) CRT inferior prognostic biomarker for guiding retreatment Schmid-Erfurth et al. (2016) IRF negatively associated with VA Schmid-Erfurth et al. (2016) SRF associated with superior visual benefits and a lower rate of progression towards atrophy blog-cacm/267674-ais-jurassic-park-moment/fulltext), we are aware of the risks that such models to propagate misinformation. Our stance is that it is easier for the human agent to verify the information in Table 2 and to annotate it with provenance data, instead of manually collecting it from literature. From the technical perspective, the burden is how to feed the GPT-3 with relevant ”prompts” (e.g. based on BioVerbNet) to get relevant information. In line with C. Baquero (https://bit.ly/3ElW1J7, prompt design was critical for querying of such language models. The job of Prompt Designer may become relevant in populating ontologies. Third, we apply reasoning to signal differences and inconsistencies among the knowledge within the ontology. These differences reflect the current understanding of the AMD disease: quantitative vs. qualitative fluid assessments, intraretinal fluid vs. subretinal fluid (SRF), exudative vs. nonexudtive fluid. For instance, for SRF both negative and positive but also no-association have been reported [2]. Moreover, heterogeneity of therapeutic approaches has been increased in the context of personalised care. This heterogeneity rises the question of inconsistent information, detected in our approach by the Racer reasoning tool. References [1] P. Mitchell, G. Liew, B. Gopinath, T. Y. Wong, Age-related macular degeneration, The Lancet 392 (2018) 1147–1159. doi:https://doi.org/10.1016/S0140- 6736(18)31550- 2 . [2] L. Kodjikian, M. Parravano, A. Clemens, , et al., Fluid as a critical biomarker in neovas- cular age-related macular degeneration management: literature review and consensus recommendations, Eye 35 (2021) 2119–2135.