Missing social history makes automated medical coding challenging

[JoakimEdin]

Prerequisites

• [ x] Put an X between the brackets on this line if you have done all of the following:
  • Checked the online documentation: https://mimic.mit.edu/
  • Checked that your issue isn't already addressed: https://github.com/MIT-LCP/mimic-code/issues?utf8=%E2%9C%93&q=

Description

Automated medical coding (also called medical code prediction) is a growing machine learning task that aims to predict medical codes given a discharge summary. MIMIC-IV has become a popular dataset to train and evaluate such models. However, there is an issue. Since your de-identification algorithm removed the social history section, certain annotated medical codes are impossible to predict. For instance, the medical codes representing whether the patient smokes (e.g., F17.210 and Z87.891) are often annotated in MIMIC-IV without being mentioned in the discharge summary. This is because of the missing social history.

The consequences of the missing section are that the models are trained on labels that are impossible to predict and are evaluated unfairly every time the necessary information would have been in the social history. Consequently, MIMIC-IV is a noisier dataset for automated medical coding than MIMIC-III (MIMIC-III contains the social history).

Is there a way to de-identify the discharge summaries without removing the social histories?

[alistairewj]

Thanks for raising this, I wasn't aware of it personally and that's an unfortunate side effect. It's open for discussion but I think the removal of the social history section remains useful for deidentification given the expansion of the dataset to ED patients. The approach could be improved by applying NER on the raw PHI note and allowing through only the medically relevant segments of the social history (alcohol use, smoking use, drug use). It's unlikely we'll get to doing that any time soon though, sorry.