We developed Naive Bayes (NB) classifier for text classification to information extraction from written text at CLEF eHealth 2018 challenge, task1. The data set used is called the CepiDC Causes of Death Corpus. It comprises French biomedical text reports of death causes. To extract ICD10 codes for each death certificate, a preprocessing process must be carried out, for example, we removed all terms from the certificates that are not related to medicine and after that we used a NB classifier to generate a classification model. The evaluation of the proposed approach does not show good performance compared with the results obtained by the other participants in the challenge.
The CLEF eHealth 2018 [
Registration of medical causes of death is mainly motivated by prevention: identify
and quantify causes deaths on which it is possible to act to reduce the avoidable
mortality [
The CLEF e-Health 2018 Task 1 CepiDC Gold Standard Training data comprises the text of 65,843 death certificates and associated gold standard ICD10 codes.
Our approach to deal with this problem is to integrate techniques of information
extraction and among of the goals of task is to foster the development of NLP tools
for French in spite of the known discrepancies in language resources available for
French and other languages in the biomedical domain [
The rest of the paper is organized as follows: Section 2 describe the process of building a classification model. Section 3 presents the formal model of the NB classifier. Section 4 presents the CLEF eHealth dataset and explains the different pretreatments performed on this dataset. Section 5 is reserved to the evaluation of our approach, it discusses the obtained results. Finally, section 5 concludes the paper. 2
The aim of the pre-processing process is to make clear the border of each language
structure and to eliminate as much as possible the language dependent factors.
Tokenization, stop word elimination and stemming are the concrete processes applied
in this step [
The documents are represented by a great amount of features and most
of them could be irrelevant or noisy [
After feature extraction, the
most important step in preprocessing the text
classification, we do Feature Selection (FS) to construct a vector space. This step
select
concise
features from the original features (
≤ ). The features can be more
and more efficient to represent the contents of the text. FS is performed
by keeping the words with highest score according to predetermined measure
of the word importance [
Naïve Bayes classifier is a simple probabilistic classifier based on applying Bayes’
Theorem with strong independence assumptions. A more descriptive term for the
underlying probability model would be independent feature model. These
independence assumptions of features make the features order is irrelevant and consequently
that the present of one feature does not affect other features in classification tasks [
Denote a vector of variables D = <di> , i=1,2,…n, represent document, where di is corresponding to a letter, a word, or other attributes about some text in reality, and a set of C ={ c1,c2,…ck } is predefined classes. Text classification is to assign a class label cj, j =1,2,..., k from C to a document [10].
Bayes classifier is a hybrid parameter probability model in essence: (1) (2) Where P(cj) is prior information of the appearing probability of class cj, P (D) is the information from
observations, which is the knowledge from the text itself to be classified, and P(D| cj) is the distribution probability of document D in classes space. Bayes classifier is to integrate this information and compute separately the posteriori of document D falling into each class cj, and assign the document to the class with the highest probability, that is [10]
Assume the components di of D are independent with each other since conditional probability P(D| cj) cannot be computed directly in practice. Thus: The model with the above assumption is called Native Bayes model, and equation (1) becomes: Because the sample information P(D) is identical to each class cj, j =1,2,..., k , equation (2) Becomes [10] : (4) (5) 4 4.1
The data set is called the CepiDC Causes of Death Corpus. It comprises free-text descriptions of causes of death as reported by physicians in the standardized causes of death forms. Each document was manually coded by experts with ICD-10 per international WHO standards. It should be noted that only one ICD10 code is provided per line. The French dataset was available in the raw and aligned formats it has about 65,843 death certificates, a set of documents in .csv format. The size of the dataset is about 27,4 Mo in compressed status and approximately 198 Mo after extracting. 4.2
We used a list of medical concepts (built by standard text as dictionary) to extract
medical concepts from the documents in the dataset by removing all their terms that
are not in the list. The reason for pruning the text using the ICD10 dictionaries is to
leave in the analyzed text only the important terms related to the treated field.We used
Weka toolkit1 to extract concept, and in the step of feature selection we used filters
methods. In order to evaluate a feature, filters use an evaluation metric that measures
the ability of the feature to differentiate each class [
The following Erreur ! Source du renvoi introuvable.3 and 4 gives a result for our model in both aligned and raw format.
Raw_all The performance of the proposed method show acceptable results for the raw format comparing with the aligned format and this is refer to the choice of good filter in the step of selection feature and according to the analysis of the classes predicted of ICD10 codes. Our system predicts one ICD10 per line. It should be noted that we intend to compare the use of the dictionary to other forms of pruning in future work. We also would like in the future to compare our proposed approach with the other methods of machine learning after we understand the domain and know how to skip and treat well the difficulties such as time and fittings.
It’s necessary to build an efficient model for information extraction system however this is not easy and still a challenge for participating groups in CLEF eHealth.
Our experiment shows that the NB classifier does not give a good result. A step of reducing the dimensionality is therefore necessary and it can improve the results. The evaluation of the proposed approach does not show good performance compared with the results obtained by the other participants in the challenge. However, in future work, we will keep finding out advanced methods in features selection to refine corpora so that they only contain suitable features, and experiment various models of machine learning for text classification. 10. McCallum, A., Nigam, K.: A comparison of event models for Naive Bayes text classification. In: Learning for Text Categorization. In: the AAAI Workshop,AAAI, pp. 41–48 Technical Report WS-98-05 (1998).