Error Propagation in EHRs via Copy/Paste: An Analysis of Relative Dates

We present a method for identifying errors in EHRs caused by copying and pasting text containing relative dates. Our method utilizes a sequence alignment method for recognizing instances of copy/paste, and regular expressions for relative dates. We furthermore present an analysis of our method on the MIMIC-II dataset. Introduction Electronic health records (EHR) are a powerful enabling technology, but concerns exist about the dangers of copying text from one clinical record to another. [1,2] While strategies exist for mitigating copy/paste in text mining approaches, little work has been done to evaluate the types of errors introduced to the EHR through copy/paste. This work studies errors due to the copying of relative dates. When a note contains a relative date (e.g., “3 days ago”, “past 4 weeks”, “43 y/o”), performing temporal reasoning requires this date be grounded, typically to the date the report was written. However, if the text is copied to a later report, the ability to ground this date is lost. We propose a method called DupLink, which links pasted text back to its original source (Figure 1). We then utilize regular expressions to identify relative dates in pasted text, enabling such cases to be flagged or even automatically corrected.