The lab events are queried from the MIMIC-III Clinical Dataset, chunked by sets of subject identifiers to monitor query progress. The dictionary of item identifiers are also queried fro mthe MIMIC-III Clinical Dataset, without chunking due to the small query size.
The LOINC database is read from a .csv file, and loaded into a dataframe.
Using a dataframe merge on the LOINC codes, the extra metadata from the LOINC database is added to the lab item dictionary. Another merge with the lab events dataframe is performed to add the LOINC metadata to the lab events.
The classification performance of the LOINC codes is summarised below. Each bar represents one of the 'classes' of metadata provided by a LOINC code. You can read more about the parts of a LOINC code on their website.
The classification performance of all non-optional LOINC code parts is 97.8%.
The top 30 values for the most salient part, the component, is plotted below.
