Deep EHR Literature Review - 1 https://qdata.github.io/deep2Read Presenter: Chao Jiang IEEE Journal of Biomedical and Health Informatics, 2017 https://qdata.github.io/deep2ReadPresenter: Chao Jiang Deep EHR Literature Review - 1 IEEE Journal of Biomedical and Health Inform / 18
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Deep EHR Literature Review - 1
https://qdata.github.io/deep2Read
Presenter: Chao Jiang
IEEE Journal of Biomedical and Health Informatics, 2017
https://qdata.github.io/deep2ReadPresenter: Chao JiangDeep EHR Literature Review - 1IEEE Journal of Biomedical and Health Informatics, 2017 1
/ 18
Outline
1 What is EHR
2 What can we do with EHR
3 Deep EHR ApplicationsEHR Information ExtractionEHR Representation LearningOutcome Prediction
4 MIMIC dataset
https://qdata.github.io/deep2ReadPresenter: Chao JiangDeep EHR Literature Review - 1IEEE Journal of Biomedical and Health Informatics, 2017 2
/ 18
Electronic Health Record (EHR)
EHR systems store data associated with each patient encounter,including demographic information, current and past diagnoses,laboratory tests and results, prescriptions, radiological images, clinicalnotes, and more.
In part due to the Health Information Technology for Economic andClinical Health (HITECH) Act of 2009, which provided $30 billion inincentives for hospitals and physician practices to adopt EHR systems.
According to the latest report from the Office of the NationalCoordinator for Health Information Technology (ONC), nearly 84% ofhospitals have adopted at least a Basic EHR system, a 9-fold increasesince 2008.
Additionally, office-based physician adoption of basic and certifiedEHRs has more than doubled, from 42% to 87%
https://qdata.github.io/deep2ReadPresenter: Chao JiangDeep EHR Literature Review - 1IEEE Journal of Biomedical and Health Informatics, 2017 3
/ 18
What can we do with EHR
While primarily designed for improving healthcare efficiency from anoperational standpoint, many studies have found secondary use for clinicalinformatics applications.
medical concept extractionpatient trajectory modelingdisease inferenceclinical decision support systems
https://qdata.github.io/deep2ReadPresenter: Chao JiangDeep EHR Literature Review - 1IEEE Journal of Biomedical and Health Informatics, 2017 4
/ 18
Deep EHR Application
https://qdata.github.io/deep2ReadPresenter: Chao JiangDeep EHR Literature Review - 1IEEE Journal of Biomedical and Health Informatics, 2017 5
/ 18
Outline
1 What is EHR
2 What can we do with EHR
3 Deep EHR ApplicationsEHR Information ExtractionEHR Representation LearningOutcome Prediction
4 MIMIC dataset
https://qdata.github.io/deep2ReadPresenter: Chao JiangDeep EHR Literature Review - 1IEEE Journal of Biomedical and Health Informatics, 2017 6
/ 18
Single Concept Extraction
Structured prediction models for RNN based sequence labeling inclinical text (EMNLP 2016)
Bidirectional RNN for Medical Event Detection in Electronic HealthRecords (NAACL 2016)
https://qdata.github.io/deep2ReadPresenter: Chao JiangDeep EHR Literature Review - 1IEEE Journal of Biomedical and Health Informatics, 2017 7
/ 18
Single Concept Extraction
Example task:The follow-up needle biopsy results were consistent with bronchiolitisobliterans, which was likely due to the Bleomycin component of his ABVDchemo. In this sentence, the true labels are Adverse Drug Event(ADE) forbronchiolitis obliterans and Drugname for ABVD chemo.
https://qdata.github.io/deep2ReadPresenter: Chao JiangDeep EHR Literature Review - 1IEEE Journal of Biomedical and Health Informatics, 2017 8
/ 18
Single Concept Extraction
https://qdata.github.io/deep2ReadPresenter: Chao JiangDeep EHR Literature Review - 1IEEE Journal of Biomedical and Health Informatics, 2017 9
/ 18
Temporal Event Extraction
Brundlefly at SemEval-2016 Task 12: Recurrent Neural Networks vs.Joint Inference for Clinical Temporal Information Extraction(SemEval-2016)
Incremental Knowledge Base Construction Using DeepDive (TheVLDB Journal)
Example Task:http://alt.qcri.org/semeval2016/task12/
They train three RNN models: a character-level RNN for tokenization; andtwo word-level RNNs for POS tagging and entity labeling.
https://qdata.github.io/deep2ReadPresenter: Chao JiangDeep EHR Literature Review - 1IEEE Journal of Biomedical and Health Informatics, 2017 10
/ 18
Relation Extraction
Clinical Relation Extraction with Deep Learning (International Journalof Hybrid Information Technology)
Example task: treatment X improves/worsens/causes condition Y, or testX reveals medical problem Y
https://qdata.github.io/deep2ReadPresenter: Chao JiangDeep EHR Literature Review - 1IEEE Journal of Biomedical and Health Informatics, 2017 11
/ 18
Outline
1 What is EHR
2 What can we do with EHR
3 Deep EHR ApplicationsEHR Information ExtractionEHR Representation LearningOutcome Prediction
4 MIMIC dataset
https://qdata.github.io/deep2ReadPresenter: Chao JiangDeep EHR Literature Review - 1IEEE Journal of Biomedical and Health Informatics, 2017 12
/ 18
Concept Learning
Distributed Embedding: word2vec, this part is very similar to learningword embeddings in NLP field.
Latent Encoding: Autoencoders (AE)
https://qdata.github.io/deep2ReadPresenter: Chao JiangDeep EHR Literature Review - 1IEEE Journal of Biomedical and Health Informatics, 2017 13
/ 18
Patient Learning
Med2Vec framework to derive distributed vector representations ofpatient sentences, i.e. ordered sequences of ICD-9, CPT, LOINC, andNational Drug Codes (NDC).
Deepr framework uses a simple word embedding layer as input to alarger CNN architecture for predicting unplanned hospital readmission.
Miotto et al. directly generate patient vectors from raw clinical codesvia stacked AEs, and show that their system achieves bettergeneralized disease prediction performance as compared to using theraw patient features.
https://qdata.github.io/deep2ReadPresenter: Chao JiangDeep EHR Literature Review - 1IEEE Journal of Biomedical and Health Informatics, 2017 14
/ 18
Patient Learning
Choi et al. derive patient vectors by first generating concept andencounter representations via skip-gram embedding, and then usingthe summed encounter vectors to represent an entire patient historyto predict the onset of heart failure.
Doctor AI system utilizes sequences of (event, time) pairs occurring ineach patients timeline across multiple admissions as input to a GRUnetwork. At each time step, the weights of the hidden units are takenas the patient representation at that point in time, from which futurepatient statuses can be modeled and predicted.
https://qdata.github.io/deep2ReadPresenter: Chao JiangDeep EHR Literature Review - 1IEEE Journal of Biomedical and Health Informatics, 2017 15
/ 18
Outline
1 What is EHR
2 What can we do with EHR
3 Deep EHR ApplicationsEHR Information ExtractionEHR Representation LearningOutcome Prediction
4 MIMIC dataset
https://qdata.github.io/deep2ReadPresenter: Chao JiangDeep EHR Literature Review - 1IEEE Journal of Biomedical and Health Informatics, 2017 16
/ 18
Scalable and accurate deep learning for electronic healthrecords (Jeff Dean @ Google Brain)
https://qdata.github.io/deep2ReadPresenter: Chao JiangDeep EHR Literature Review - 1IEEE Journal of Biomedical and Health Informatics, 2017 17
/ 18
MIMIC-III
The latest version of MIMIC is MIMIC-III v1.4, which comprises over58,000 hospital admissions for 38,645 adults and 7,875 neonates. Thedata spans June 2001 - October 2012.
https://qdata.github.io/deep2ReadPresenter: Chao JiangDeep EHR Literature Review - 1IEEE Journal of Biomedical and Health Informatics, 2017 18
/ 18
Related Documents
