Guest Post: New techniques to improve computer-assisted coding accuracy

by Erica E. Remer, MD, FACEP, CCDS

Computer-assisted coding (CAC) is to native, manual coding as reading Dickens’ Great Expectations for an English class is to reading the book for pleasure (in my opinion). Having to analyze the theme, character development, setting, symbolism, and language affords you a completely different experience than just enjoying the story. Coding depends on understanding the story of the patient encounter and being able to coherently translate it into accurate, specific codes.

The best CAC users first read the record and understands the sequence of events and how the encounter unfolded. The coder can then can examine each offered code and assess its applicability. In fact, the most difficult part of using CAC is resisting the urge to relinquish the overall responsibility of coding to the computer.

Novices and experienced coders alike may be overwhelmed by the large number of autosuggested codes—for example, one superuser I interviewed says that CAC responses to a medical record for a patient with a 12-day stay offered him 55 codes to choose from. This makes it quite challenging to separate the wheat from the chaff. Additionally, much of the CAC color-coding can be quite distracting; your eyes are drawn from one highlighted word/phrase to the next, and you might be disinclined to read the plain text snippets between—and those  could contain important information.

Sometimes, in an attempt to make the electronic health record (EHR) more functional, providers alter the font of their text. Attending physicians bold (or italicize or use a different color, etc.) their additions to a resident’s or advanced practice practitioner’s note. CAC does not take formatting into consideration, so the attending physician’s alterations hold no more sway than anyone else’s. This may set up conflicts, or a coder may miss an attending discounting a resident’s diagnosis.

Content versus context

Computers are much better at assessing content than they are at context, but coding often requires command of both. For example, the computer may take verbiage of “heart attack” and autosuggest code I25.2 (old myocardial infarction), even if the doctor qualifies it with “last week.” If the coder doesn’t notice the fact that the myocardial infarction was within 28 days, they may lose a legitimate MCC.

A fact of clinical life is that sometimes patients get admitted for persistent or severe signs or symptoms, and no definitive diagnosis is ascertained after study; patients get better despite us. Sometimes, however, CAC may suppress a sign or symptom if it finds a definitive condition manifested by that sign or symptom (e.g., weakness/acute kidney injury [AKI]). If the only diagnosis a coder has for a four-day stay is AKI which resolved on day two, the patient ends up erroneously in the renal failure DRG. DRG 684, Renal failure without CC/MCC, may have more favorable metrics, but DRG 948, Signs and symptoms without MCC, would be the correct DRG.

Another example is the symptom of agitation or restlessness getting suppressed by the accepted CAC suggestion of F32.9 (major depressive disorder, single episode, unspecified) for the verbiage of “history of depression.” Agitation and restlessness are sometimes accompanied by behavioral disturbance in dementia, but the coder needs to see the big picture to tease that out.

One system shared with me a case where a patient had chronic kidney disease (CKD) with superimposed AKI requiring emergent dialysis. The CAC suggested end-stage renal disease (ESRD) as a diagnosis and excluded the AKI and CKD. This is a tough concept to explain to a machine –dialysis is usually, but not exclusively, associated with the chronic condition of ESRD.

Editor’s note: This article originally appeared in JustCoding. Dr. Remer is the founder and president of Erica Remer, MD, Inc., Consulting Services. Contact her at icd10md@outlook.com. Opinions expressed are that of the author and do not necessarily represent HCPro, ACDIS, or any of its subsidiaries.