THIS CODE IS LICENSED UNDER THE APACHE PUBLIC LICENSE (APL) 2.0
This is Dr. Herskovic's medication reconciliation algorithm. It takes two lists of medications and returns three lists: a reconciled list, the part of list 1 that could not be reconciled, and the part of list 2 that could not be reconciled.
This implementation requires RXNorm, which is part of the UMLS. You can extract RXNorm from the UMLS to use with this program by using the included generate_rxnorm_file.py script. You must have an RRF version of the UMLS to use this script. Invoke it as:
python generate_rxnorm_file.py /path/to/umls/metathesaurus rxnorm.pickle
This might have funny results if run using the OS X System python, which is missing the bsddb library that the shelve module uses. A 64-bit Linux or a python version from python.org on a 64-bit OS X are highly recommended. The script generates several "shelve" files containing UMLS concepts, relationships, etc., all with the suffix you passed as the last argument. All of these files are necessary to run the reconcile script.
The algorithm can also use a treatment database in a pickled bz2 file. You may ignore this file, and treatment intent reconciliation will be skipped. If you want to provide one, it should be a pickle dictionary of the form
{'CUI_of_drug': set(['CUI_of_condition', 'CUI_of_another_condition', etc.])}
I apologize in advance, but I can not share the dataset we used to create our copy of the treatment file, nor the treatment file itself. If you want your own copy, please contact me and I'll do my best to put you in touch with the people who own the dataset. The algorithm will still work without it.
After you have your rxnorm pickled file, please try out:
python reconcile.py
to perform a demo reconciliation. You'll need all of the files that end in .rxnorm.pickle (or whatever name you gave it) for it to work.
I also included the SMARTApp version of the reconciliation algorithm. You will need Josh Mandel's smart client modules, available from https://github.com/chb/smart_client_python and placed in the smart_client subdirectory (and, of course, all its dependencies). Try:
python reconcile_smart.py 8000
and run it inside the SMART sandbox to play with it.
Our original evaluation of this algorithm was published in the Proceedings of the AMIA Symposium: Bozzo Silva PA, Bernstam EV, Markowitz E, Johnson TR, Zhang J, Herskovic JR. Automated medication reconciliation and complexity of care transitions. Proc AMIA Ann Symp 2011.
The JSON output contains 5 key-value pairs: original_list_1, original_list_2, new_list_1, new_list_2, and reconciled. original_list_1 and 2 echo the input to the program. reconciled contains the merged list. new_list_1 and new_list_2 contain the items from original_list_1 and 2 that could not be reconciled.
The reconciled list is a list of key-value pair stores. Each item in the list has at least four key-value pairs: med1, score, identical, and mechanism.
mechanism describes the way in which medications were merged (if they were identical, or their ingredients matched, etc.) identical contains a refinement of mechanism, and specifies which fields matched between two drugs. score gives the quality of the match (between 0 and 1, with 1 being complete certainty) med1 contains a single medication. If mechanism is not "identical", there will be a med2 as well; if mechanism is "identical", you can assume that the second medication was identical to the first one, and is therefore not provided.
med1 contains:
- id: This is a sequential id generated arbitrarily by the software that uniquely identifies a medication from its input. Every medication the program sees in a single run has a different id; ids will repeat between runs.
- medication_name: This is the medication's name.
- dose: The medication's dose (i.e. 100)
- units: The units of the medication's dose (i.e. mg)
- formulation: The physical format of the medication (tablet, capsule, etc)
- instructions: Whatever free-form instructions were provided for the medication.
- original_string: The original input string
- provenance: Whatever provenance was specified in the input
- normalized_dose: The dose in a normalized format for a day ("W XYZ", W being dose of an individual unit, X being units, Y being the number of times per day (i.e. frequency), and Z is the number of units to be delivered simultaneously. So, for example, 2 500 mg Aspirin tid becomes 500 mg32)
- frequency: The number of times per day the drug is administered.
Route is currently not available.
med2 obviously has the same format as med1.