Text applications in healthcare subsection + image (greenelab#567)

content/03.categorize.md

@@ -88,7 +88,7 @@ of certified dermatologists [@tag:Codella2016_ensemble_melanoma;
 Reusing features from natural images is also an emerging approach for radiographic images,
 where datasets are often too small to train large deep neural networks without
 these techniques [@tag:Bar2015_nonmed_tl; @tag:Shin2016_cad_tl;
-@tag:Rajkomar2017_radiographs; @tag:Lakhani2017_radiography]. 
+@tag:Rajkomar2017_radiographs; @tag:Lakhani2017_radiography].
 A deep CNN trained on natural images
 boosts performance in radiographic images [@tag:Rajkomar2017_radiographs]. However, the target task required
 either re-training the initial model from scratch with special pre-processing or
@@ -209,7 +209,7 @@ demonstrated that NegBio is highly accurate for detecting negative and equivocal
 findings (~90% in F-measure, which balances precision and recall
 [@doi:10.1038/nmeth.3945]). These highly-accurate results meet the need to
 generate a corpus with weak labels, which serves as a solid foundation for the
-later process of image classification. The resulting dataset 
+later process of image classification. The resulting dataset
 [@url:https://nihcc.app.box.com/v/ChestXray-NIHCC] consists of 112,120
 frontal-view chest X-ray images from 30,805 patients, and each image is
 associated with one or more weakly-labeled pathology category (e.g. pneumonia
@@ -231,6 +231,62 @@ a need for better pathologist-computer interaction techniques that will allow
 combining the power of deep learning methods with human expertise and lead to
 better-informed decisions for patient treatment and care.
 
+### Text applications in healthcare
+
+Due to the rapid growth of scholarly publications and electronic medical/health records, biomedical text mining has become increasingly important in recent years.
+The main tasks in biological and clinical text mining include, but are not limited to, named entity recognition, relation/event extraction, and information retrieval (Figure @fig:biotm).
+
+![Deep learning applications, tasks and models based on NLP perspectives.](images/biotm.png){#fig:biotm width="100%"}
+
+Machine learning is widely used in many of aforementioned tasks where engineering the optimal set of features is a must.
+However it is often difficult to know what features should be extracted, even with domain knowledge.
+Deep learning addresses this challenge by enabling "computers to build complex
+concepts out of simpler ones [@tag:goodfellow2016deep]".
+Moreover, deep learning has shown competitive performance compared to traditional ML methods.
+As a result, we have observed growing interests of using deep learning in biomedical text mining applications in recent years.
+In this section, we summarize recent studies in text mining with various deep learning methods.
+We categorize the body of work by the application domain/text genre (biomedical literature vs. clinical notes) and by the actual task (e.g. concept or relation extraction).
+
+In biomedical text mining, named entity recognition (NER) is a task of identifying text spans that refer to a biological concept of a specific class, such as disease or chemical, in a controlled vocabulary or ontology.
+NER is of importance as it is often needed as a first step in many complex text mining systems.
+The current state-of-the-art methods typically reformulate the task as a sequence labeling problem and use conditional random fields [@doi:10.1093/bioinformatics/btw343; @doi:10.1093/bioinformatics/btt156; @doi:10.1093/bioinformatics/btt474].
+In recent years, word embeddings that contain rich latent semantic information of words have been widely used to improve the NER performance.
+Liu et al. studied the effect of word embeddings on drug name recognition and compared them with traditional semantic features [@doi:10.3390/info6040848].
+Tang et al. investigated word embeddings in gene, DNA, and cell line mention detection tasks [@doi:10.1155/2014/240403].
+Moreover, Wu et al. examined the use of neural word embeddings for clinical abbreviation disambiguation [@doi:10.18653/v1/w15-3822].
+Liu et al. exploited task-oriented resources to learn word embeddings for clinical abbreviation expansion [@doi:10.18653/v1/w15-3810].
+
+Relation extraction is a task of detecting and classifying semantic relationship between entities from the literature.
+At present, kernel methods or feature-based approaches are commonly applied [@doi:10.1371/journal.pcbi.1000837; @doi:10.1186/s13321-016-0165-z; @doi:10.1093/bioinformatics/btp602].
+There are several proposals of using deep learning in the biomedical relation tasks because it can relieve the feature sparsity and engineering problems.
+Some studies focused on jointly extracting biomedical entities as well as their relations simultaneously [@tag:li2016joint; @doi:10.1186/s12859-017-1609-9], while others applied deep learning on relation classification with given entities.
+For example, Peng and Lu proposed a multichannel dependency-based CNN and Hua and Quan proposed a shortest path based CNN for the sentence-based protein-protein extraction task [@doi:10.18653/v1/w17-2304; @doi:10.1155/2016/8479587; @doi:10.1155/2016/1850404].
+Jiang et al. proposed a biomedical domain-specific word embedding model to reduce the manual labor of designing semantic representation for the same task [@doi:10.1504/IJDMB.2016.074878].
+Gu et al. employed a maximum entropy model and a CNN model for chemical-induced disease relation extraction at inter- and intra-sentence level, respectively [@doi:10.1093/database/bax024].
+For drug-drug interaction, Zhao et al. used a CNN that employs word embeddings with the syntactic information of a sentence as well as features of POS tags and dependency trees [@doi:10.1093/bioinformatics/btw486].
+Masaki et al. experimented with an attention CNN, and Yi et al. a recurrent neural network model (RNN) model with multiple attention layers [@doi:10.18653/v1/w17-2302; @arxiv:1705.03261].
+In both cases, it is a single model with attention mechanism, which allows the decoder to "attend" to different parts of the source sentence. As a result, it does not require dependency parsing or training multiple models. 
+Both attention CNN and RNN have comparable results, but the CNN model has an advantage in that it can be easily computed in parallel, hence making it faster with recent Graphical Processing Units (GPUs).
+
+For biotopes event extraction, Li et al. employed CNN and distributed representation while Mehryary et al. used long-short term memory (LSTM) networks to extract complicate relations [@doi:10.18653/v1/w16-3012; @doi:10.18653/v1/w16-3009].
+Li et al. applied word embedding to extract complete events from biomedical text and archived the results comparable to the state-of-the-art systems [@doi:10.18653/v1/w15-3814].
+There are also some works that identified event triggers rather than the complete event [@doi:10.1142/S0219720015410012;  @arxiv:1705.09516].
+Taken together, it appears that deep learning models outperform traditional kernel methods or feature-based approaches by 1-5% in f-score.
+Among various deep learning approaches, CNN stands out as the most popular model both in terms of computational complexity and performance, while RNN has achieved continuous progress.
+
+Information retrieval is a task of finding relevant text that satisfies an information need from within a large document collection.
+While deep learning has not yet achieved the same level of success in this area as seen in others, the recent surge of interest and work suggest that this may be quickly changing.
+For example, Mohan et al. described a deep learning approach to modeling the relevance of a document's text to a query, applied to the entire biomedical literature [@doi:10.18653/v1/w17-2328].
+
+In the clinical domain, Jagannatha and Yu employed a bidirectional LSTM RNN structure to extract adverse drug events from electronic health records [@pmid:27885364].
+Choi et al. developed Doctor AI, a RNN-based model that can learn efficient patient representation from a large amount of longitudinal patient records and predict diagnosis and medication code of EHR [@arxiv:1511.05942].
+Minarro-Giménez et al. applied the word2vec [@arxiv:1301.3781] deep learning toolkit to medical corpora and evaluated the efficiency of word2vec in identifying properties of pharmaceuticals based on mid-sized, unstructured medical text corpora without any additional background knowledge [@doi:10.3233/978-1-61499-432-9-584].
+Lin et al. investigated using CNN to extract temporal relations [@doi:10.18653/v1/w17-2341].
+Karimi et al. investigated the applicability of deep learning at autocoding of radiology reports using the International Classification of Diseases [@doi:10.18653/v1/w17-2342; @tag:world2004international].
+
+To summarize, deep learning has shown promising results in many biomedical text mining tasks and applications.
+But to further realize its full potential in this domain, either large size of labeled data and/or technical advancements in current methods coping with limited labeled data are required.
+
 ### Electronic health records
 
 EHR data include substantial amounts of free text, which remains challenging to

content/citation-tags.tsv

@@ -264,3 +264,6 @@ Zhou2015_deep_sea	doi:10.1038/nmeth.3547
 Zhu2016_advers_mamm	doi:10.1101/095786
 Zhu2016_mult_inst_mamm	doi:10.1101/095794
 Zintgraf2017_visualizing	arxiv:1702.04595
+goodfellow2016deep	url:http://www.deeplearningbook.org/
+li2016joint	url:http://www.ijcai.org/Proceedings/16/Papers/403.pdf
+world2004international	url:http://www.who.int/classifications/icd/en/