ICD 10 CM Codes Seem Incorrect.

[vanessailana]

Hello, I am conducting a study on the amazing ICD-10 Codes from the Physionet dataset. I am examining a subset of the clinical notes, and I have noticed that some notes do not mention the ICD diagnosis associated with them. For instance a patient could be associated to the hypothermia ICD-10 CM code, but there is no mention of hypothermia in any of the clinical notes associated to them.

How were these codes derived?

Best,
Vanessa

[alistairewj]

Can you give the hadm_id for your example?

The codes are more or less directly from the source EHR. These are billed after hospital discharge, and codes are determined by reviewing signed notes from healthcare providers. It is possible that it is an incorrect billing code, or it is possible that we did not include the relevant note that mentions the diagnosis.

[vanessailana]

If it is an incorrect billing code, does it mean the dataset is reliable for ICD Code Prediction? This has happened in maybe 20 notes and then https://eicu-crd.mit.edu/ dataset. If you would like, I can show you the results I found.

[vanessailana]

@alistairewj For example, for ham_id=21037483, the only ICD-10 CM code I see associated with it is G4.3801. However, when I read the note, I see the individual has other diagnoses like:

Past Medical History:
Complex migraines (this was covered)
Asthma
Medullary kidney cysts
Strabismus s/p surgery
Vertigo/vertiginous migraine

[alistairewj]

I should have been clearer - it's always possible for there to be a few errors (they will happen if you look at 100k+ hospitalizations). However, it should be rare. If you start to see a systematic issue that's when I'd start to worry that we made a mistake in the build. One ICD code for a hospitalization seems low, so I'll take a look. I will say these are only the hospital billed codes; we don't have any provider billing data in MIMIC.

[Anaudia]

Hey, we are currently facing a similar problem. We tried to use the MIMIC IV dataset to train a model for ICD-10-CM coding. However, we quickly realized that some of the codes in the dataset do not have corresponding information in the discharge summaries. To date, we have evaluated a few hundred discharge summaries and concluded that information is missing most often for the following codes:

icd_code	missing_count	count	missing_proportion
I471	23	23	1.000000
I482	22	22	1.000000
M545	21	21	1.000000
I272	20	20	1.000000
I472	17	17	1.000000
G20	12	12	1.000000
Z9114	28	28	1.000000
E872	65	65	1.000000
R740	32	32	1.000000
G92	40	40	1.000000
N183	58	60	0.966667
R51	19	20	0.950000
T814XXA	11	12	0.916667
Z23	26	40	0.650000
Z87891	151	316	0.477848
F17210	46	100	0.460000
Z006	13	30	0.433333
N400	25	63	0.396825
Y929	35	90	0.388889
Y92239	19	49	0.387755
Y838	15	42	0.357143

We believe some of these codes, such as Z87891, may be related to an issue that was previously raised. If you are interested, I can provide you with all the hadm_ids and the corresponding codes for which information is missing.