Discharge Summaries

Why discharge is the hardest documentation moment

Discharge summaries serve two distinct readers and have to satisfy both. The receiving primary-care or specialty provider needs a complete clinical picture: hospital course, problems addressed, medication changes with rationale, pending labs, follow-up plan, and red flags. The patient leaving the hospital needs the same information at a sixth-grade reading level, in plain language, with the medications and the warning signs clear enough to act on alone at home.

Both readers consistently get failed. Discharge instructions exceed the recommended reading level for ~88.7% of internal-medicine patients. Medication non-reconciliation is identified in ~50% of inpatient episodes, with 63% of those non-reconciled events carrying potential for moderate harm and 2% for severe harm. Production delays in discharge summaries are independently associated with readmission rates. AI drafting addresses every part of this — readability, comprehensiveness, medication clarity, time-to-completion — when the workflow is designed around clinician review and explicit medication reconciliation.

The dual-reader pattern that works

The reference architecture in the 2025–2026 literature converges on a two-output workflow: one clinician-grade summary for the receiving provider, one patient-friendly version for the patient. Same source material, same medication list, same plan — different language and structure.

What the published evidence shows

• checkQuality parity on the clinician summary. The Dutch academic-hospital evaluation (292 paired summaries, eight blinded clinicians) found mean Likert quality scores of 3.67 (LLM) vs 3.77 (physician) — comparable. LLM drafts were more concise and coherent; physician drafts were more comprehensive.
• checkPatient comprehension improves materially. NEJM AI's GPT-4 plain-language translation study found subjective comprehension scores rose 2.4 points and objective scores rose 1.2 points across diagnoses — with the largest gains in populations with historically low health literacy.
• checkReadability shifts. Expert evaluators rated AI-generated patient-friendly summaries comprehensible in 88–97% of cases — versus the baseline where ~88.7% of original instructions exceed sixth-grade reading level.
• checkMultimodal extensions. Stanford's ED-Explain (PSB 2026) showed AI-generated discharge instructions were significantly more complete, correct, and accessible than originals — with personalized video as an additional surface for low-literacy patients.
• closeThe error count caveat. The same Dutch study reported 2.91 errors per LLM summary vs 1.82 per physician summary. Lower-comprehensiveness errors and fact-omissions are the failure modes. This is the reason "review and sign" must be a real review step, not a one-click confirmation.

Where it goes wrong — the patterns to plan around

• closeMedication-list errors. The dominant safety failure. 88% of patients have at least one medication discrepancy after discharge (median 3 per patient). Common patterns: unintentional discontinuation of a chronic medication, inappropriate retention of an inpatient-only medication, dose / frequency drift, missing rationale for a change. Mitigation: explicit reconciliation block, enforced before signature.
• closeComprehensiveness drops. LLM drafts read better but skip more. Mitigation: structured input (the chart artifacts the LLM is given) must include problem list, med list, recent labs, pending tasks, and active orders — not just the encounter narrative.
• closePlain-language overshoot. Aggressive simplification can drop clinically critical nuance. Mitigation: keep the clinician-grade summary as ground truth, derive the patient version from it, and have the same clinician review both.
• closeLanguage and equity gaps. Comprehension gains are largest where health literacy is lowest — but accuracy in non-English drafts depends on model coverage. Evaluate per language, not in aggregate.
• close"Sign without reading" automation bias. Pre-filled draft is harder to scrutinize than a blank one. Mitigation: friction in the review UI (require explicit acknowledgement of the medication block), random spot audit by pharmacy or a second clinician, edit-distance tracking.

The evaluation rubric that survives the demo

Cloud commercial vs. on-prem — the architecture choice

Discharge drafting touches the broadest patient-data surface in the hospital: full hospital course, complete medication history, problem list, labs, vitals, plan. The cloud-vs-on-prem choice is therefore most often decided by which architecture can credibly take that surface inside the inference path.

For most hospitals, the on-prem path is the cleaner architecture for discharge specifically — because the data surface is largest, the safety stakes are highest, and the medication reconciliation step benefits most from being hosted next to the pharmacy system rather than across a vendor boundary.

How this fits into a multi-app local stack

Discharge drafting shares infrastructure with the rest of the on-prem clinical AI stack: the same retrieval layer that backs document Q&A sources patient-instruction language from the approved patient-education library; the same audit log that captures policy lookups captures discharge edits; the same FHIR integration that grounds ambient scribes grounds discharge medication reconciliation; the same handoff-style summary that feeds shift handoffs gets reused for inter-facility transfer.

Pick a unit, run a real pilot

The fastest path to a defensible discharge-summary deployment is to scope one unit (typically internal medicine or hospital medicine), define the rubric above, and run a 60-day pilot with mandatory clinician review and explicit medication-reconciliation tracking. Compare a cloud commercial draft tool against an on-prem reference stack on every metric — but particularly on medication accuracy, time-to-summary, and patient comprehension at teach-back.

send Request a WalledCare pilot arrow_back Back to directory