nlp2fhir-deep-learning

Integration of NLP2FHIR Representation with Deep Learning Models for EHR Phenotyping

Prerequsites

• Processed texts as a FHIR Resource Bundle from NLP2FHIR
• Python 3.5 and above
• Tensorflow >= 1.4.0
• fhirclient
• pandas
• networkx
• nltk

Dataset preparation

Input files

For both the i2b2 and mimic datasets, the input file structure under data/obesity_datasets/{$DATASET}/ (replace {$DATASET} to i2b2 or mimic) should be as follows:

• Notes:
  • i2b2: obesity notes/REPORT{$NOTE_ID}.txt
  • mimic: notes/{$NOTE_ID}.txt
• Resource Bundle (the output directory of NLP2FHIR): ``
  • i2b2: ObesityResourceResourceBundle/REPORT{$NOTE_ID}.txt.json
  • mimic: ObesityResourceResourceBundle/{$NOTE_ID}.txt.json
• Gold standard labels
  • i2b2: train_groundtruth.xml and test_groundtruth.xml
  • mimic: a csv file with columns of ${NOTE_ID} and all comobidities as 0 and 1, and the last 2 columns as the indicator of training and testing, also in 0 and 1.

Run

Run python fhir_{$DATASET}_reader.py

Run python remove_words.py {$DATASET}

Run python build_graph.py {$DATASET}

Run python train_on_keras_all_com.py

References

The implementation of Text GCN :

• Liang Yao, Chengsheng Mao, Yuan Luo. "Graph Convolutional Networks for Text Classification." In 33rd AAAI Conference on Artificial Intelligence (AAAI-19), 7370-7377 [Paper] [Code]
  • Require: Python 2.7 or 3.6, Tensorflow >= 1.4.0

For FHIR-EHR cases

• Hong et al., Developing a FHIR-based EHR phenotyping framework: A case study for identification of patients with obesity and multiple comorbidities from discharge summaries, JBI [Paper]