Computational Healthcare Library

This repository contains Computational Healthcare library (chlib), the underlying library used in Computational Healthcare. Computational Healthcare library is designed to allow computer scientists to use large healthcare claims databases. Using chlib you can easily load & process large healthcare databases with millions of patients.

Analyze up to 200 Million visits & 70 Million patients

Currently we support following three databases:

1. Texas Inpatient public use data file 2006-2009: This database contains approximately 11 Million de-identified inpatient visits from Texas during 2006-2009, this data is available as free download. Inpatients visits in this database lack patient identifier.

2. HCUP Nationwide Readmission database for 2013: This database contains de-identified inpatients visits during 2013 & 14. Unlike Texas database all inpatient visits are associated with a patient identifier and its possible to track patient through multiple visits.

3. HCUP State Inpatient, ED & SASD database: This is one of the largest logitudinal database of medical claims in the world. Acquiring this database typically takes several weeks and can cost few 100$ to ~10,000$ depending number of states/years/types. We currently support data from California, Florida & New York. If you are interested in using Computational Healthcare with this dataset please contact us.

Please note that this repository does not contains any data, nor do we provide any data. You should acquire the datasets on your own from AHRQ and/or other state agencies.

Architecture & Data Model

A quick overview of data model, architecture is available in this presentation.

• The library uses Protocol buffers for storing raw data (visits & patients) and aggregate statistics.
• Categorical fields-values are represented as enums using Protocol buffers.
• Raw data is stored in a levelDB database
• Protocol Buffers and LevelDB makes it easy to use any programming language
• We provide code to compute aggregate statistics in privacy preserving manner
• Integrated with TensorFlow for building machine learning models

Installation & Setup

• The docker folder contains a Dockerfile with all dependencies specified. It also contains script for building docker image, starting container, preparing databases from user supplied files.

• Once you have obtained data you should modify docker/prepare_nrd.sh script with correct path to .CSV file and run the script.

• For Texas dataset modify/run the prepare_tx.sh.

• The dockerfile uses TensorFlow version 0.11 docker image as a starting point. Thus in addition to Computational Healthcare library it also contains a Jupyter notebook server which runs automatically when the container is started. Once the preparation step is complete you can use the jupyter notebook server running (inside the container) on port 8888 of your local machine.

Quick overview

    import chlib
    NRD = chlib.data.Data.get_from_config('../config.json','HCUPNRD')
    # patients
    for p_key,patient in NRD.iter_patients():
        break
    print p_key,patient
    
    # visits
    for p_key,patient in NRD.iter_patients():
        for v in patient.visits:
            break
        break
    print v

output (fake)

123213213 patient_key: "213213213"
visits {
  key: "123213"
  patient_key: "213123213213"
  dataset: "NRD_2011"
  state: "NRD"
  facility: "1232131"
  vtype: IP
  age: 23
  sex: FEMALE
  race: R_UNKNOWN
  source: S_ED
  disposition: D_ROUTINE
  los: 0
  death: ALIVE
  payer: PRIVATE
  primary_diagnosis: "D80"
  primary_procedure {
    pcode: "P86"
    pday: 0
    ctype: ICD
  }
  drg: "DG05"
  prs {
    pcode: "P8659"
    pday: -1
    ctype: ICD
    occur: 1
  }
  year: 2013
  month: 10
  quarter: 1
  zip: Z_THIRD
  dnr: DNR_UNAVAILABLE
  charge: 229.0
}
raw: "<>"
linked: true

  key: "123213"
  patient_key: "213213213123"
  dataset: "NRD_2011"
  state: "NRD"
  facility: "1232131"
  vtype: IP
  age: 65
  sex: FEMALE
  race: R_UNKNOWN
  source: S_ED
  disposition: D_ROUTINE
  los: 0
  death: ALIVE
  payer: PRIVATE
  primary_diagnosis: "D80"
  primary_procedure {
    pcode: "P86"
    pday: 0
    ctype: ICD
  }
  drg: "DG05"
  dxs: "D709"
  prs {
    pcode: "P8659"
    pday: -1
    ctype: ICD
    occur: 1
  }
  year: 2013
  day: 5151
  month: 10
  quarter: 1
  zip: Z_THIRD
  dnr: DNR_UNAVAILABLE
  charge: 229.0

Text description of codes and enums

    coder =  chlib.codes.Coder() 
    print 'D486',coder['D486'] # ICD-9 diagnosis codes are prepended with 'D'
    print 'P9971',coder['P9971'] # ICD-9 procedure codes are prepended with 'P'
    print coder[chlib.entity.enums.D_AMA] # You can also print string representation of Enums            

output

D486 Pneumonia, organism unspecified
P9971 Therapeutic plasmapheresis
Against medical advice

Retrieve list of patients with particular diagnosis or procedure

    patients_undergoing_plasmapheresis = [p for _,p in NRD.iter_patients_by_code('P9971')]
    # You can speed this up by precomputing list of patients for each codes, using 'fab precompute'
    print len(patients_undergoing_plasmapheresis)
    for v in patients_undergoing_plasmapheresis[0].visits:
        print v.key,v.day,v.prs

Tutorials / Articles

1. Computational Healthcare for reproducible machine learning: building embedding from million inpatient visits

2. Analyzing long term outcomes of Ventriculostomy in pediatric patients

3. Exploring OHDSI common data model, comparison with Computational Healthcare (currently writing)

Issues & Bugs

To minimize chances of visit/patient level information leaking via Exceptions messages or Traceback, we have not enabled issues on github repo. If you find any bugs, make sure that your bug report/question does not contains any visit or patient level information. To file a bug please email us at address provided below.

Contact

For more information, comments or if you plan on citing Computational Healthcare library please contact Akshay Bhat at [email protected].

Copyright

Copyright Cornell University 2016; All rights reserved; Please contact us for more information.