Grado de magíster en informática médica / Actualmente los informes radiológicos se redactan en texto libre sin un campo específico que los categorice según diagnóstico. Por este motivo, la identificación de los diagnósticos clasificados como patología crítica debe hacerse de forma manual, acarreando consigo problemas como el submuestreo y gran tiempo invertido. Este trabajo propone como solución desarrollar una herramienta utilizando métodos de procesamiento de lenguaje natural para analizar los texto de forma masiva. En esta tesis se plantea como hipótesis que es posible identificar más del 80% de los diagnósticos existentes en SNOMED-CT (una terminología médica) presentes en las impresiones de los informes radiológicos, identificando la patología crítica con más de un 90% de sensibilidad mediante algoritmos de procesamiento de lenguaje natural (NLP). Para clasificar los informes se utilizó SNOMED-CT por su amplio manejo de conceptos médicos y sinónimos. La tarea se realizó con 3 algoritmos: 1) un motor de búsqueda para encontrar los términos de SNOMED-CT contenidos en los informes utilizando indexación reversa, 2) un detector de negación basado en expresiones regulares y 3) se combinó ambas herramientas para identificar patología crítica. Los algoritmos propuestos fueron evaluados en muestra representativa (n=219) de 1973 informes de Angiografía Pulmonar por Tomografía Computada, etiquetada por 2 médicos. Como resultados se obtuvo un valor kappa de acuerdo entre etiquetadores de 85.5%, IC95%[80.8-90.3%], p < 0.001. Por otra parte el motor de búsqueda presentó un rendimiento con medida F (F) de 0.94, sensibilidad (S) de 91.2% y valor predictivo positivo (VPP) de 98%. El detector de negación obtuvo una F de 0.99, S de 98.7% y VPP de 99.3%. Para medir el rendimiento en la detección de patología crítica se utilizó como referencia el diagnóstico de tromboembolismo pulmonar (TEP), obteniendo valores F de 0.94, S de 96.3% y VPP de 92.86% Como conclusión, el presente trabajo de tesis muestra que es posible construir una herramienta para identificar la patología crítica basada en NLP utilizando la regularidad de los patrones de expresión en el texto, lo que permitirá en futuros trabajos crear herramientas de soporte para la toma de decisiones. / Currently radiology reports are written in free text without a specific field to categorize according to diagnosis. Therefore, identification of diagnostics listed as critical result, group characterized by having a high risk of harm to the patient, must be done manually. As a solution is proposed the use of natural language processing tools to analyze big volume of texts. This thesis pose the hypothesis that it is possible to identify more than 80% of existing diagnostics from impressions of radiology reports on SNOMED-CT, a clinical terminology, identifying critical results with more than 90% sensitivity, using natural language processing (NLP) algorithms. To identify reports, SNOMED was used because of its wide management of medical terms and synonyms. Identification was built as a 3 steps algorithm: 1) A search engine was built to find terms of SNOMED contained in reports using reverse indexing, 2) a negation detector based on regular expressions, and 3) both tools were combined to identify critical results. The proposed algorithms were tested against a representative sample (n = 219) of 1973 Computed Tomography Pulmonary Angiography (CTPA) reports, which were tagged by 2 medical doctors. The obtained results were an inter-rater reliability kappa value of 85.5% for taggers, was obtained IC95% [80.8-90.3%]. Moreover, search engine had a performance of measure F (F) of 0.94, sensitivity (S) of 91.2% and positive predictive value (PPV) of 98%. The negation detector had a F of 0.99, S of 98.7% and VPP of 99.3%. The measurement of performance for critical results detection was made using pulmonary embolism as reference, obtaining values; F of 0.94, S of 96.3% and VPP of 92.86% In conclusion, this thesis shows that it is possible to build a tool to identify critical results using NLP by making use of the specific regularity of text expressions in the case of radiology reports, allowing in future researchs to create decision support tools. / 2021

Two broad themes of this research are, 1) to develop a generalized framework for studying the process of patient information acquisition and 2) to develop and evaluate automated techniques for identifying domain-specific vocabulary terms contained in, or missing from, a standardized controlled medical vocabulary with emphasis on those terms necessary for representing the canine physical examination. A generalized framework for studying the process of patient information acquisition is addressed by the Patient Information Acquisition Model (PIAM). PIAM illustrates the decision-to-perception chain which links a clinician's decision to collect information, either personally or through another, with the perception of the resulting information. PIAM serves as a framework for a systematic approach to identifying causes of missing or inaccurate information. The vocabulary studies in this research were conducted using free-text with two objectives in mind, 1) develop and evaluate automated techniques for identifying canine physical examination terms contained in the Systematized Nomenclature of Medicine and Veterinary Medicine (SNOMED), version 3.3 and 2) develop and evaluate automated techniques for identifying canine physical examination terms not documented in the 1997 release of the Unified Medical Language System (UMLS). Two lexical matching techniques for identifying SNOMED concepts contained in free-text were evaluated, 1) lexical matching using SNOMED version 3.3 terms alone and 2) Metathesaurus-enhanced lexical matching. Metathesaurus-enhanced lexical matching utilized non-SNOMED terms from the source vocabularies of the Metathesaurus of the Unified Medical Language System to identify SNOMED concepts in free-text using links among synonymous terms contained in the Metathesaurus. Explicit synonym disagreement between the Metathesaurus and its source vocabularies was identified during the Metathesaurus-enhanced lexical matching studies. Explicit synonym disagreement occurs, 1) when terms within a single concept group in a source vocabulary are mapped to multiple Metathesaurus concepts, and 2) when terms from multiple concept groups in a source vocabulary are mapped to a single Metathesaurus concept. Five causes of explicit synonym disagreement between a source vocabulary and the Metathesaurus were identified in this research, 1) errors within a source vocabulary, 2) errors within the Metathesaurus, 3) errors in mapping between the Metathesaurus and a source vocabulary, 4) systematic differences in vocabulary management between the Metathesaurus and a source vocabulary, and 5) differences regarding synonymy among domain experts, based on perspective or context. Three approaches to reconciling differences among domain experts are proposed. First, document which terms are involved. Second, provide a mechanism for selecting either vocabulary-based or Metathesaurus-based synonymy. Third, assign a "basis of synonymy" attribute to each set of synonymous terms in order to identify the perspective or context of synonymy explicitly. The second objective, identifying canine physical examination terms not documented in the 1997 release of the UMLS was accomplished using lexical matching, domain-specific free-text, the Metathesaurus and the SPECIALIST Lexicon. Terms contained in the Metathesaurus and SPECIALIST Lexicon were removed from free-text and the remaining character strings were presented to domain experts along with the original sections of text for manual review. / Ph. D.