Medical Terminology in Patient Medical Records - Public Data Sets

Question

I am interested in sample data of real patient medical records (anonymized or demographics removed completely) for the purpose of running through NLP system - specifically diagnoses, admissions and progress notes - anything where medical terminology is used. I have no interest in linking data to specific patient types or categories - I just want the text of the notes

Ideally I would like a data such as:

10/4/2011 - 59 year old male admitted with right sided chest pain, reduced air entry and fever for 7 days. X-ray showed occlusion of right lower lobe consistent with community-acquired bacterial pneumonia.Non-smoker.

13/4/2011 - Reviewed by MO. Left leg swelling noted with some tenderness on extension - DVT suspected. No fever, SOB. BP 120/90. Ix to follow. U/S ordered. Heparin commenced.

Are there any free data sets with this sort of data? I have looked at various public data repositories such as http://www.healthdata.gov/ and https://www.nlm.nih.gov/hsrinfo/datasites.html without much luck.

Answer 1

The MIMIC-III dataset would be well suited to the kind of natural language processing (NLP) study that you are interested in doing. MIMIC-III includes deidentified nursing progress notes, imaging reports, and discharge summaries for tens of thousands of patients who were admitted to intensive care units in the US.

You can read about several NLP studies that have used the MIMIC dataset on PubMed (note that most use MIMIC-II, an earlier version of MIMIC). For example, the following study explored text classification methods for use in risk adjustment: http://www.ncbi.nlm.nih.gov/pmc/articles/PMC4147615/

While MIMIC-III is a freely available dataset, the information that it contains is sensitive and so there is a formal process for requesting access. For details, see: http://mimic.physionet.org/gettingstarted/access/

Answer 2

Another good-size data set of natural language patient notes is available here: http://www.mtsamples.com/

It covers notes across multiple specialities and conditions.

We can also provide an analytics-ready (CSV/JSON) version of this dataset. Please contact me if that's of interest.

Good luck!