LLM-Based Clinical Report Summarisation

Problem

Discharge summarisation was a high-friction step in hospital throughput — clinicians spent significant time consolidating records before patients could be released, slowing both care and bed turnover.

Approach

1. 1.Designed an LLM pipeline that ingests structured EHR fields and unstructured notes, applies clinical-context prompts, and produces a structured discharge summary draft.
2. 2.Built an evaluation pipeline aligned with clinical standards — covering factuality, omission, and policy compliance — reviewed by clinicians for sign-off.
3. 3.Applied hallucination mitigation via output grounding, schema enforcement, and few-shot exemplars curated by clinical reviewers.
4. 4.Integrated into clinician workflow as a 'draft, then review' assistive tool — humans always in the loop.

Learnings

• In regulated domains the evaluation pipeline matters more than the model. Trust is earned on the eval suite.
• Designing for assistive rather than autonomous use was the unlock for clinician adoption.