AB1767-HPR DOCUMENT SEARCH IN LARGE RHEUMATOLOGY DATABASES: ADVANCED KEYWORD QUERIES TO SELECT HOMOGENEOUS PHENOTYPES - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2023

AB1767-HPR DOCUMENT SEARCH IN LARGE RHEUMATOLOGY DATABASES: ADVANCED KEYWORD QUERIES TO SELECT HOMOGENEOUS PHENOTYPES

Résumé

Background Natural language processing tools are powerful for mining rheumatology databases, extracting patient information directly from clinical notes. However, these algorithms come with a high computational cost and are often not applicable at the scale of very large databases in the temporality of clinical practice. Objectives The objective of our study is the automatic detection of clinical documents of interest for a specific clinical question, with low computational cost, to be applied on a database of millions of documents. These sets of documents of interest constitute a pre-screening to allow the development of more complex algorithms. Methods The task was considered as an information retrieval task in French clinical texts. Two different methods were compared. For the first method, we used several state-of-the-art document vector representations: TF-IDF, doc2vec, docBERT and tested if the closest documents are relevant. The second method consists in building a powerful query expansion from a key term entered, its French synonyms from the UMLS and the synonyms found by similarity with the embeddings of the CODER algorithm. These methods are developed and evaluated on a set of 8 and on 20 phenotypes respectively (e.g. “pericarditis in lupus”, etc.). Our database corresponds to 2 million documents from a cohort of patients suffering from four autoimmune diseases: systemic lupus erythematosus, scleroderma, antiphospholipid syndrome, and Takayasu’s disease, coming from the AP-HP’s data warehouse. Results Our experience does not support the vector representation model of clinical notes for searching similar patients. However, searching with an advanced synonym search method can lead to very good results without additional burden for the clinician: we achieved a precision (or positive predictive value) of 0.93 [0.90; 0.96] evaluated manually by a physician and a recall (or sensitivity) of 0.78 [0.71; 0.85] evaluated on the basis of the ICD10 codes of the retrieved patients. Conclusion We propose a new advanced keyword search method with automatic synonym search with very good accuracy and recall performance. References [1]Alison Callahan, Vladimir Polony, José D Posada, Juan M Banda, Saurabh Gombar, Nigam H Shah, ACE: the Advanced Cohort Engine for searching longitudinal patient records, Journal of the American Medical Informatics Association , Volume 28, Issue 7, July 2021, Pages 1468–1479, [2]Yuan, Zheng, et al. “CODER: Knowledge-infused cross-lingual medical term embedding for term normalization.” Journal of biomedical informatics 126 (2022): 103983 [3]Gérardin C, Mageau A, Mékinian A, Tannier X, Carrat F, Construction of Cohorts of Similar Patients From Automatic Extraction of Medical Concepts: Phenotype Extraction Study, JMIR Med Inform 2022;10(12):e42379 Table 1. Accuracy and recall results for 13 over 20 queries. Query Accuracy (on 50 manually-annotated document per query) Recall (comparison with respective CIM10) Number of corresponding documents 1 “Rheumatoid Arthritis” 0.98 0.73 15189 2 “Takayasu” 1 0.94 2459 3 “Pericarditis in lupus” 0.92 0.93 7490 4 “Kidney transplantation” 0.92 0.98 10716 5 “Autoimmune hepatitis” 0.8 0.85 2797 6 “Dermatomyositis” 1.0 0.77 3510 7 “Idiopathic thrombocytopenic purpura” 0.98 0.81 3749 8 “Acute kidney injury” 0.86 0.81 15775 9 “Raynaud syndrome” 0.98 0.98 31900 10 “HIV” 0.90 0.98 43582 11 “Scleroderma” 1.0 0.92 24199 12 “Diabetes” 0.96 0.96 51224 13 … “Stroke” … 0.64 0.63 28162 Overall 0.93 [0.90; 0.96] 0.78 [0.71; 0.85] Figure 1. Overview of the two methods of searching for documents in our data warehouse. Method 1 is document oriented and method 2 is keyword oriented. Acknowledgements The authors would like to thank the AP-HP data warehouse, which provided the data and the computing power to carry out this study under good conditions. We would like to thank all the medical colleges, including internal medicine, rheumatology, dermatology, nephrology, pneumology, hepato-gastroenterology, hematology, endocrinology, gynecology, infectiology, cardiology, oncology, emergency and intensive care units, that gave their agreements for the use of the clinical data. Disclosure of Interests None Declared.

Dates et versions

hal-04137383 , version 1 (22-06-2023)

Identifiants

Citer

C. Gérardin, Y. Xong, A. Mekinian, F. Carrat, Xavier Tannier. AB1767-HPR DOCUMENT SEARCH IN LARGE RHEUMATOLOGY DATABASES: ADVANCED KEYWORD QUERIES TO SELECT HOMOGENEOUS PHENOTYPES. European Alliance of Associations for Rheumatology, May 2023, Milan, Italy. pp.2117.1-2118, ⟨10.1136/annrheumdis-2023-eular.6231⟩. ⟨hal-04137383⟩
2 Consultations
0 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More