DISCHARGE SUMMARY
ملخص الخروجAn AI tool that drafts the discharge summary in about 90 seconds, with a source behind every line and the clinician still the author.
In pilot since 2026. The screens shown include the refinements from the pilot's first months.
SUMMARIES TYPED FROM SCRATCH, AT THE END OF A SHIFT
وش المشكلة؟Discharge summaries were typed from scratch in ICIS or pasted from earlier notes. They took 30 to 45 minutes, read differently from one clinician to the next, and held up discharge. The brief was an AI that writes. The job was trust: a clinician signs every word.
EVERY LINE HAS A SOURCE
كل سطر له مصدرEach section can be traced back to the notes it came from, in one click, with the used text highlighted.
BLANK BEATS WRONG
الفراغ أحسن من الغلطIf the notes don't say it, the summary says so. The AI never fills a gap with a guess.
THE CLINICIAN IS THE AUTHOR
الطبيب هو الكاتبThe AI drafts. The clinician chooses the inputs, edits any section, and signs. Nothing leaves without them.
WHAT CLINICIANS TOLD US
وش لقيتI watched summaries being written before designing one, then tested the riskiest idea first: trust.
| METHOD | WITH WHOM | SAMPLE |
|---|---|---|
| Shadowing | Residents and physicians on discharge days | 8 discharge rounds |
| Interviews | Physicians, residents and medical records | 16 people |
| Summary audit | Past discharge summaries | 60 summaries |
| Prototype tests | Clinicians with real-shaped data | 2 rounds of 7 |
1 in 3 past summaries had a section copied from an earlier stay
Audit of 60 discharge summaries written before the tool
Copied: a section matching an earlier admission word for word.
NO TRACE, NO TRUST
Every clinician in testing opened a source before accepting a diagnosis.
NOT EVERY NOTE MATTERS
Admissions pile up repeated notes. A few carry the whole story.
A BLANK BEATS A GUESS
A confident wrong answer is the fastest way to lose a clinician.
THREE DECISIONS THAT SHAPED IT
وش قررتA SOURCE BEHIND EVERY LINE
Each section opens its sources in one click, with the text the summary used highlighted. Clinicians check a diagnosis in seconds instead of rereading every note.
THE CLINICIAN CHOOSES THE NOTES
Feeding all 73 notes to the model was slower and worse. The clinician picks from grouped notes, so they own the inputs.
11 of 73 notes held what the summary needed
Every note on the test encounter; green squares are the ones the clinician chose
Most unchosen notes repeat information from the main notes.
HONEST WHEN IT DOESN'T KNOW
If the notes don't cover a section, the summary says so and offers to add text or choose more notes. Nothing goes to ICIS until the clinician confirms they're the author, with every gap listed.
THE FULL JOURNEY
الرحلة كاملة
WHAT CHANGED IN THE PILOT
وش تغيّرMeasured from tool logs, send confirmations and a pilot survey.
A sendable summary in 8 minutes, not 35
Median time from choosing notes to sending to ICIS
Before: typed in ICIS or pasted from earlier notes. Pilot: tool session logs.
Every section is traceable
Share of summaries in the pilot
From tool logs and send confirmations.
“I still read every line, but now I'm reading, not typing.”
Resident, NeurosciencesWHAT I BROUGHT BEYOND THE SCREENS
كيف مشّيت المشروعTurned 'an AI that writes summaries' into three trust principles with clinical stakeholders.
Put View source in the very first prototype, before any visual design.
Defined what 'author' means with the business analyst and stakeholders, and built it into every step.
WHAT I LEARNED
وش تعلّمتAgree the section list with clinicians before the model is built, and log edits per section from day one. Edits are the clearest signal of where the AI is wrong.