From Hours of Manual Work to Clinician-Ready ECG Reports – Generated in Minutes, Without Leaving Their Network

Brief

We built a fully offline AI reporting pipeline for Curtin University that takes ECG classification outputs and generates complete, visually rich PDF reports in minutes. Powered by a local installation of Google’s Gemma 3.0, the system runs entirely on Curtin’s infrastructure, so no patient data leaves the network, no external APIs are involved, and no cloud costs accumulate. The result is faster research, more consistent reporting, and a system that scales without introducing privacy risk.

Very impressed with your work. Thanks for all your help to date!

Dr Andrew Maiorana, Curtin University

Problem

ECG classification models can identify rhythm categories with strong accuracy, but a classification output on its own isn’t a clinical report, it’s a starting point. Turning that output into something a clinician can actually read and act on requires interpreting confidence scores, contextualising lead-by-lead feature data, summarising patient demographics, and writing clear narrative text across multiple sections. For Curtin’s research team, that process was manual, time-consuming, and required significant clinical knowledge to execute well. It was also a bottleneck: every new ECG added to the dataset meant more report-writing work before the research could progress. They needed a system that could handle the translation from model output to clinical document automatically, and do it in a way that kept every piece of patient data firmly inside their own infrastructure.

Solution

We delivered a complete AI-powered ECG reporting pipeline, hosted entirely on Curtin’s own hardware. The system ingests three inputs: ECG classification outputs across seven rhythm categories, detailed lead-by-lead clinical feature data, and patient demographics, and produces a structured, styled PDF report ready for clinical review. Each report includes narrative-style interpretations with diagnosis confidence scores, highlighted feature abnormalities for quick triage, lead-specific time-series plots, visual classification probability charts, and a patient demographic summary.

• Fully on-premises: Gemma 3.0 runs on Curtin’s own hardware, so no patient data ever reaches an external server, cloud platform, or third-party API
• Multi-input synthesis: the system combines classification outputs, granular clinical features, and patient context into a single coherent document, not just a data dump
• Clinically structured output: reports are formatted for how clinicians actually read them, narrative first, supporting data clearly laid out, abnormalities flagged for fast triage
• Consistent at scale: every report follows the same structure and standard, regardless of volume, with no variation from clinician to clinician
• Built on open-weight infrastructure: using Gemma 3.0 means no licensing costs and full control over the model deployment

Result

Curtin’s research team can now generate detailed, clinician-ready ECG reports in minutes, down from hours of manual work per report. The quality is consistent, the format is immediately useful, and the system handles the full translation from raw model output to readable clinical document without any human intervention in between. For a research programme working at scale, that’s not just a time saving, it’s what makes the research tractable. The team can process more ECGs, iterate faster, and spend their time on analysis and clinical judgement rather than report formatting. The system also sets a strong foundation for future development, as the reporting pipeline can be updated to match as the dataset grows and classification categories evolve, without rebuilding from scratch.

Under the Hood