An AI-powered quality assurance pipeline for hospital discharge summaries — ensuring every patient crosses from hospital to community care with a complete, accurate, and clinically safe handover document.
A discharge summary is the final clinical document produced before a patient leaves hospital. It carries a patient's entire hospital story — diagnoses, treatments, medications, follow-up plans — to the community care team who will look after them next. When it is incomplete or inaccurate, that information gap can cause real patient harm.
Every discharge summary carries a patient's full hospital story to the team who will care for them next. When it's incomplete or inaccurate, that gap can cause real harm.
Studies show that 25–60% of discharge summaries contain at least one clinically significant omission or error — yet most are only reviewed after the patient has already left. Horizon changes that.
Horizon is an AI-powered QA platform that evaluates every discharge summary before it reaches the GP — checking for completeness, verifying factual consistency against the patient record, and flagging any content that doesn't match the source data.
The result: a scored, graded report that gives clinical teams a clear, actionable picture of handover quality, every time.
Submit - Upload a discharge summary (text or PDF) alongside the patient record (FHIR JSON or plain text) via the web interface or REST API.
Completeness check - Horizon evaluates the summary against a structured clinical question set, scoring each component as full, partial, or missing. Medication changes and follow-up plans are weighted more heavily than administrative details.
Consistency check - Key factual claims in the summary are verified against the patient record. Any content that can't be substantiated is flagged as a hallucination and penalised in the final score.
Score and report - A weighted quality score and letter grade (A–D) are generated instantly, along with a downloadable PDF report detailing findings across every component.
Each component is evaluated as full, partial, or missing based on structured questions. Weighted scores reflect clinical priority — medication changes and follow-up plans carry more weight than administrative details.
Claude extracts 6–8 specific factual claims from the summary and verifies each against the patient record. Fabricated content is reported separately as hallucinations with a score penalty.
Completeness is weighted more heavily as the primary clinical concern. Hallucinations carry a fixed per-item penalty capped at 30 points.
A ≥ 85 — High quality
B ≥ 70 — Good quality
C ≥ 55 — Needs attention
D < 55 — Poor quality
Horizon ships with three built-in question sets, each reflecting a different evaluation context. Sets are fully customisable — clinicians can add, remove, rename and reweight components to match local guidelines and specialty requirements.
Universal baseline covering admission details, diagnoses, procedures, test results, medications, follow-up and administrative completeness. Suitable for any inpatient specialty.
Built directly from the SA Health Handover to GP Document Guide. Evaluates SBAR structure, medication change documentation, investigation appropriateness, and patient-accessible language.
Tailored for cardiac admissions. Includes LVEF documentation, ECG and echocardiogram results, anticoagulation details, cardiac intervention outcomes, and cognitive and social assessment.
Create specialty-specific sets for oncology, aged care, mental health, paediatrics, or any local clinical guideline. Sets are saved to the browser and exported via the API for institution-wide deployment.
Patient record text and discharge summary text are sent to the configured AI provider — either the Anthropic API or Azure AI Foundry — for evaluation. PDF extraction happens client-side; the PDF file itself is never transmitted. When using Azure AI Foundry in Australian regions, data remains within Australian data boundaries.
Question sets, evaluation history, PDF reports, and provider settings (including API keys and Azure endpoint configuration) are stored entirely in the browser. No evaluation results are stored on any server. When using the Python API backend, the AI provider key is held server-side and never exposed to the browser.
Horizon routes all clinical data to a dedicated, private model deployment owned and operated by Healthful AI. This means:
Important: Horizon is a quality improvement tool, not a clinical decision support system. Evaluation results should be reviewed by a qualified clinician. Final responsibility for discharge summary quality rests with the treating team.
Copyright © 2025 Healthful AI Pty Ltd - All Rights Reserved.