Healthcare
Healthcare Administration
The problem
The assistant drafts the post-visit summary in minutes, but the diagnosis ends up in the message subject line, the doctor's proposed next steps turn into clinical interpretation, or the report goes to the patient as a message attachment instead of through the portal.
30 courses · 5 specific to this profession
A process from your work
How long does it take you to code the diagnoses on a discharge summary?
- Step 1The tool reads the doctor's text and suggests diagnosis codes.
- Step 2You check each suggestion against the official list, then accept, change or reject it.
- Step 3When the text is unclear, you ask the doctor; the tool does not fill in the code.
- Where the tool goes wrongThe tool picks the unspecified femur fracture code, though the diagnosis specifies the neck.
- The verification stepCompare the code's official description word by word with the doctor's diagnosis.
More: what was measured and the full check
What changed, as measured
For longer clinical notes, median coding time fell by 123 seconds, a 46% reduction; for shorter notes, the difference of 11 seconds was not statistically significant.
Who was measured: Nurses, coding experts and physicians in Sweden and Norway, coding clinical notes with and without the AI suggestions in a crossover randomised trial.
The caveat: A very small sample, and the tool covered only one chapter of the classification (diseases of the digestive system).
The check, step by step
- Open the official ICD-10 list (the WHO browser or the list your institution uses) and look up the code the tool gave. Read the official description, not the description the tool wrote.
- Compare the official description word by word with the diagnosis on the discharge sheet. The words "unspecified" or "other" in the code, when the diagnosis is specific, mean the code is wrong.
- Look at the other codes in the same category and choose the one that matches the doctor's text. If none matches or the diagnosis is not clearly legible, ask the doctor and do not fill it in yourself.
- Note next to the line who checked the code and against which list. The line is not sent without this note.
The course covers summaries, tables and reports made with AI, each with its source stated and a check before any decision.
We can help you with this process: “Healthcare administration AI: Data and analytics”, 8 hours.
A documented failure pattern
S72.9 instead of S72.0
An administrator in the hospital records office asks an AI tool for the ICD-10 code of the diagnosis the doctor wrote on the discharge sheet, to enter it in the reporting and billing line for the service.
Source: Icahn School of Medicine at Mount Sinai, 2024, and 3 more.
Why it is easy to miss
The doctor's diagnosis specifies the part of the femur, the neck, which has its own code, S72.0; the tool proposes S72.9, which in the official WHO list means a fracture of the femur with the part unspecified. It has dropped exactly the word that makes the code right. Next to the code it puts the words "fracture of femur", taken from the diagnosis itself, so the eye sees a match, and the code exists in the right category, S72, so no format check rejects it. In the Mount Sinai study, models often returned codes that were close in meaning but more general than the precise code.
Where the failure pattern is documented: Academic studies: NEJM AI (Icahn School of Medicine at Mount Sinai, April 2024), a medRxiv preprint by the same group (July 2023), an arXiv preprint (July 2024); the World Health Organization's official ICD-10 classification (2019 version).Sources: Icahn School of Medicine at Mount Sinai, 2024; medRxiv, 2023; arXiv, 2024; World Health Organization, 2019.The case above is ours, written for work in Albania; the laws it names were checked against the official text.
What it looks like
Discharge sheet, the doctor's diagnosisFracture of the neck of the femur.
The tool's proposed line"ICD-10 code: S72.9, fracture of femur."
Translated from the Albanian original.
Specialist courses for this profession
Healthcare administration AI: Orientation
Healthcare administration AI: Workflow
Healthcare administration AI: Documents and communication
Healthcare administration AI: Data and analytics
Healthcare administration AI: Quality and evaluation
What you leave with
The reception card lists the warning signs for post-visit summaries (a diagnosis in the subject line, clinical interpretation, an attachment instead of the portal), and for each one the source and the role that confirms it. It is used before anything goes to a patient, including by a new colleague.
Healthcare Administration
The catalogue lists all 30 together, with filters for format and time.