Bio NLP Information Extraction from Radiology Reports Emilia

Bio. NLP, Information Extraction from Radiology Reports Emilia Apostolova College of Computing and Digital Media De. Paul University

Bio. NLP – conferences and shared tasks Pacific Symposium on Biocomputing Intelligent Systems for Molecular Biology Association for Computational Linguistics North American Association for Computational Linguistics Bio. NLP Bio. Creative TREC Genomics IClef

Information Extraction (in Bio. Medicine) The NLP Pipeline • • • Lexical Analysis – tokenization, morphological analysis, linguistic lexicons. Syntactic Analysis – Part of Speech Tagging, Chunking, Parsing. Semantic Analysis – Lexical Semantic Interpretation, Semantic Interpretation of Utterances.

NLP Pipeline Frameworks • • GATE - General Architecture for Text Engineering. Apache UIMA - Unstructured Information Management Application. Geneways - a system for automatically extracting, analyzing, visualizing and integrating molecular pathway data from the research literature. PASTA - Protein Structures and Information Extraction from Biological Texts.

Lexical Analysis - Tokenization Segmenting text into linguistic tokens – words and sentences. • • • Abbreviations - The Study was conducted within the U. S. Apostrophes - IL-10's cytokine synthesis inhibitory activity Hyphenation - co-operate, cooperate Multiple formats: 464, 285. 23 and 464295. 23 Sentence boundary detection - : , ; , -

Lexical Analysis – Morphological analysis Link surface variants of a lexical element to its canonical base form. E. g. inflections (activat-es, activat-ed, activat-ing), derivations (activation). Porter stemmer – lexicon-free approach. Finds longest match of a word to a a list of English derivational and inflectional suffixes. Two-level morphology – a finite state based approach that applies a series of parallel transducers to input tokens. (fly -> flies)

Syntactic Level – Part of Speech Tagging activation – POS noun, singular activate – POS verb, present non-3 d person singular active – POS adjective report?

Syntactic Level - Parsing A natural language parser is a program that works out the grammatical structure of sentences, for instance, which groups of words go together (as "phrases") and which words are the subject or object of a verb. The Stanford Dependency Parser - a Java implementation of probabilistic natural language parsers, trained on the Penn Treebank.

Semantic Level – Lexical Interpretation • Selectional Restrictions: transitive verbs: inhibit [something], transcribe [something] semantic restrictions: inhibit [Process], transcribe [Nucleic Acid] Syntactically admissible, but semantically invalid: to inhibit amino acids to transcribe cell growth

Discourse Level - Pragmatics • Discourse referents; what entities does a given message refer to? • What background knowledge is needed to understand a given message? • How do the beliefs of speaker and hearer interact in the interpretation of a message? • What is a relevant answer to a given question? • Summarization, Translation, Dialog Systems, Natural Language Generation.

Lexical resources for (Bio)NLP • Princeton Wordnet • NLM UMLS lexicon and metathesaurus. • The Open Biomedical Ontologies

Text and Image Integration

Automatic Image Annotation

Automatic Image Annotation Where? Woman (Population Group), Right breast (Body Part, Organ, or Organ Component) How? Mammography (Diagnostic Procedure) What? Calcification (Pathologic Function), Lesion (Finding), Carcinoma, Papillary (Neoplastic Process)

IE from Clinical Texts – Radiology and Pathology Reports Northwestern University Medical School Department of Radiology Imaging Informatics

Radiology Reports

Sample Radiology Report Patient Name: XXXXXXX, XXXXX Medical Record Number: XXXXX DOB: XXXX. XX Sex: F Accession Number: XXXX Study Requested: DIG MAMMOGRAM SCREENING (3300000) Scheduled Date and Time: XXXX. XX 13: 02: 00. 0000 Requesting Physician: XXXXXXX, Reason for Exam: V 76. 12 --------------Radiological Report----------------Comparison is made to previous exams dated XX/XX/XX. CLINICAL HISTORY: Seventy-two year old woman for screening exam. Patient has a family history of breast cancer, sister age sixty years old. Patient has a history of a previous left breast benign biopsy. TECHNIQUE: Mammograms were obtained using digital technique. FINDINGS: There is dense fibroglandular tissue bilaterally. No dominant masses or clustered microcalcifications suggestive of malignancy are seen. 1. NO SPECIFIC FEATURES OF MALIGNANCY SEEN EITHER BREAST. 2. NO SIGNIFICANT CHANGE WHEN COMPARED WITH PRIOR STUDIES. 3. ANNUAL SCREENING MAMMOGRAM IS RECOMMENDED. CODE (1): NEGATIVE Attending Radiologist: XXXXXXX, MD Date Signed off: XXXXXX, Transc. by: NS

NLP for Clinical Texts • Document retrieval – case finding. • Subject recruitment – identify patients that can benefit from a study. • Surveillance – monitoring disease outbreaks. • Discovery of disease-drug associations. • Discovery of disease-finding associations.

IE from Radiology Reports Automatic Section Segmentation Demographics History Comparison Technique Findings Impression Recommendation Sign off

Dataset 215, 000 free-text radiology reports selected randomly from 3 million reports over period of 9 years and representing 24 different types of diagnostic procedures.

Method – Training Set • Hand-crafted rules for automatic extraction of a training set. Common boundary patterns: e. g. section Findings – text between known section headers and another known headings: ^ (finding | observation | discussion)s? : ^ (W*)(finding | observation | discussion)s? (W*)$ • 3, 000 automatically segmented “highconfidence” radiology reports, containing all 8 sections of interest.

Method • Classification task - each sentence from a radiology report is assigned to one of 8 predefined report sections.

• Sentence features used for classifier. training a Sentence Orthography Possible orthographic types are All Capitals, Mixed Case, or presence of a Header pattern, such as a phrase at the beginning of a line followed by a colon. Previous Sentence Boundary Formatting boundary separating the current and previous text sentences. Possible values are white space containing new lines, white space without new lines, nonalphabetic characters, or the beginning of the file. Following Sentence Boundary Formatting boundary separating the current and next text sentences. Possible values are white space containing new lines, white space without new lines, non-alphabetic characters, or the end of the file. Cosine Vector Distance from the current sentence to each of the eight sections' word vectors. Exact Header Match This feature specifies if the sentence contains a header identified as belonging to one of the sections in the training data.

Work in Progress • Identify named entities within sections using a controlled vocabulary – findings, diseases, observations, anatomical organs, imaging modalities. • Negation Discovery. • Identify relationships between named entities of interest, for example what observations are associated with a diagnosis. • Use radiology report text to support automatic annotation of medical images.

Q/A