Our system (Team: HCMUS) combined rule-based and machine learning methods. The first step in which the test files were normalized and preprocessed. The pre-processing was related to the problems as: the special characters (dot in the case of abbreviation, ?, etc.), replacing the names and the dates in the brackets ([]). The document was split into the sections and paragraphs. Then the NLP tools were used for sentence splitting, POS tagging and parsing. The set of rules based on the dependence graph which were used to recognize events. In order to recognize the concepts (the 8th attribute), the UMLS and MetaMap were used. For the 9th attribute, the machine learning method was based on the features such as: document types, section types, temporal expressions (ago, today, etc.), explicit dates in the sentences and verb POS tags. For task 2a, this system achieved an overall accuracy of 0.827, F1score of 0.389, precision of 0.367 and recall of 0.415. For task 2b, the system performed with an F1-core, precision and recall of 0.420, 0.378 and 0.472 respectively, in the strict mode and 0.648, 0.583 and 0.729 respectively, in the relaxed mode.
ShARe/CLEFeHealth 2013 Lab offered the shared tasks: identification and
normalization of disorders and normalization of abbreviations and acronyms in the clinical
reports with respect to the terminology standards in the healthcare as well as the
information retrieval to address the questions that the patients may have while reading
clinical reports [1]. This year, ShARe/C
In this paper, we present our approach for Task 2a and 2b of the
ShARe/C
Pre-Processing
Processing Document: Some punctuations (?, -, etc.) are important triggers for
determining an attribute value. For example: If “-” and “?” stand in front of a
disease/disorder, they determine the first attribute value is “yes” and the third attribute
value is “yes” for the disease/disorder. In the dependence graph, these punctuations
did not appear, so we had to replace them with the other punctuations that suit our
system. Because medical data are sensitive and private, such as: all names of
patients, doctors and hospitals, .etc., the data are encoded and marked by some
special characters, for example: "She was transferred to [**Hospital1 27**] per
recommendation of her GI specialist Dr. [**First Name (STitle) 5060**]". In addition,
there are date time phrases which have been marked with special characters by
annotators, for instance: "She was discharged home on [**2011-02-02**]". All
special characters (like: [,*) and encoded names (like: "First Name (Stile) 5060",
"Hospital1 27", etc.) will lead to incorrect parsing. Therefore, we cleaned the data
by replacing encoded names with pseudo names and deleting all special characters.
For example, we replaced the encoded name phrase "[**First Name (Stile) 5060]"
with "Peter".
Section Splitter: Clinical notes can be considered as semi-structured data which are
split into distinct sections like CHIEF COMPLAINT, HISTORY OF PRESENT
ILLNESS, PAST MEDICAL HISTORY, PHYSICAL EXAMINATION, etc. Each
section tends to describe events of a particular timeframe. For example,
„HISTORY OF PRESENT ILLNESS‟ predominantly describes events occurring
before DOCTIME, whereas „MEDICATIONS‟ provides a snapshot at DOCTIME
and „ONGOING CARE ORDERS‟ discusses events which have not yet occurred
[6]. Some statistical analyses on the corpus show that 94% of Diseases/Disorders in
the section "PHYSICAL EXAMINATION" are OVERLAP, 90% of
Disease/Disorder in the section "CHIEF COMPLAINT" are BEFORE_OVERLAPS
and 100% of Diseases/Disorders in the section "YOU SHOULD CONTACT
YOUR MD IF YOU EXPERIENCE" are AFTER. The problem is how to split
1 http://c
CXR ([**2020-05-31**])
IMPRESSION: Scarring versus atelectasis in right lung base. No acute process. Sentence Splitter, POS tag and Parser: In this stage, natural language processing (NLP) is applied. It includes splitting sentences, tagging parts of speech, and deep parsing sentences. 2.2
Rule-based approach
This approach was applied for 9 attributes (1-10, except for 9), we used the output of the pre-processing step in order to extract the trigger sets and rule sets from the training data. The trigger sets are cue slot values corresponding to attributes of disease/disorder. Each attribute of disease/disorder has a specific trigger list. Then we enriched the trigger list by given resources. Particularly, we added triggers that have negative meaning from NegEx to the 1st attribute‟s trigger list. The set of rules is built manually based on linguistic information and dependency graph. The Stanford typed dependencies representation was designed to provide a simple description of the grammatical relationships in a sentence that can be easily understood and effectively used by the people who do not have linguistic expertise want to extract textual relations [2]. Fig.1. gives the representation of the dependency graph for an example sentence “Mitral stenosis is not present and definite mitral regurgitation is not seen.” where {not present, not seen} are triggers and {Mitral stenosis, mitral regurgitation} are diseases/disorders in the sentence.
Next step, we built the rules based on the representation of dependency graph with the aim to identify the relation between a disease/disorder and a trigger. In that system, there are two rule types used: 1) Type 1 rule is a disease/disorder which is directly relevant to a trigger and 2) Type 2 rule is a disease/disorder which is indirectly relevant to a trigger through another disease/disorder. Each attribute has its own trigger list. The rules are performed as follows: For example, the rule set identifies 1st attribute values in Fig.2. as follows: ({relation = “conj_or”}{governor = “clubbing”}{dependent = “cyanosis”}) (“cyanosis”: Yes) Or
({relation = “conj_or”}{governor = “clubbing”}{dependent = “edema”}) (“edema”: Yes)
({relation = “neg”}{governor = “clubbing”}{dependent = “No”}) (“clubbing”: Yes) Document type feature: document type that the Disease/Disorder appears. Section feature: as described in the pre-processing phase, a section is a feature to identify relation between Disease/Disorder and DocTime. Temporal expression: There are temporal expressions that help to predict the output, such as "ago", "today", "at present", etc. Explicit date feature: Explicit dates are date strings that are explicitly annotated in the clinical notes by the annotators. They can be identified by using regular expression. However, it is necessary to determine the scope of the explicit date. We divided the explicit date into two levels: 1) Sentence scope and 2) Paragraph scope. For example: disease/disorder “headache” in the sentence “In [**2015-01-14**], the patient had headache …” has explicit date (2015-01-14) in Sentence scope. An explicit date having the Paragraph scope means that this explicit date may link to all disease/disorder within the paragraph. In the experiments, we found the explicit dates in the Paragraph scope which means usually occur in the PERTINENT RESULTS section. The relation between this explicit date and the admission date, the discharge date is used as a feature for our classifier.
Verb POS tags: based on the parser, we identify verbs which link to the Disease/Disorder and their POS tags. 3
Our system used the resources such as Stanford NLP tool2, NegEx project3, MetaMap tool4, LibSVM5, Weka tool6 and UMLS7. We used the Stanford NLP tools for 2 http://nlp.stanford.edu/index.shtml 3 https://code.google.com/p/negex/ 4 http://metamap.nlm.nih.gov/
Negation Indicator (NI) Subject Class (SC) Uncertainty Indicator (UI) Course Class (CC) Severity Class (SV) Conditional Class (CO) Generic Class (GC) Body Location (BL)
DocTime Class (DT) 10
Temporal Expression (TE)
Average pre-processing texts, WEKA for classifying the 9th attribute, and MetaMap for identifying candidate UMLS concepts. MetaMap creates its final UMLS concept mapping by choosing appropriate candidates that cover as much of the input text as possible. 4
For the attributes from 1 to 7, the rule-based approach performs well. The rules based on the dependence graph have contributed to the high performance. The highest accuracy among the predicted attributes is 0.995 for the subject class (Table 1). The machine learning used to predict the 9th attribute did not produce the high performance. The reason may be due to the feature set which is not good enough to recognize the document time.
For Task 2b, our system achieves an F1-score of 0.420 in the strict evaluation (Table 2) and F1-score of 0.648 in the relaxed evaluation (Table 3) respectively, which suggests a rule-based approach that can not identify an exact cue span representing an attribute value. 5 http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/ 6 http://www.cs.waikato.ac.nz/ml/weka/ 7 http://www.nlm.nih.gov/research/umls/
We applied the rule-based approach for all attributes, except for the 9th attribute which was processed by the machine learning approach. In Task 2a, our system achieved an overall accuracy of 0.827, F1-score of 0.389, precision of 0.367 and recall of 0.415. In Task 2b, our system performed with an F1-core, precision and recall of 0.420, 0.378 and 0.472 respectively, in the strict mode and 0.648, 0.583 and 0.729, respectively, in the relaxed mode. Further improvements will be likely to be feasible by adding new features to the machine learning model and normalization of rule set, as well as adding an approach that combines the rule-based and machine learning for all attributes. These tasks will be the focus of interest in future work.