Description:
Cardiologists routinely analyze heterogeneous data sources such as ECG tracings, Doppler echocardiography reports, and electronic health records (EHRs). Integrating these inputs into a coherent clinical interpretation requires significant expertise and careful cross-referencing across multiple systems and formats.
This project aims to develop a fully offline prototype software system based on MedGemma, designed to support the integrated analysis of multimodal cardiac data. The system will transform ECG signals, echocardiographic findings, and relevant EHR data into concise and structured clinical assessments. It will generate explainable diagnostic reports that clearly document the evidence supporting each output, allowing clinicians to easily trace conclusions back to specific signal features, extracted parameters, or contextual patient information.
The prototype will follow a privacy-by-design architecture. All processing will occur in memory, without persistence of sensitive patient data. Optional network isolation will ensure that no data leaves the local environment, and logging will be limited strictly to anonymized performance metrics.
The current scope focuses on technical prototyping and engineering validation using public datasets. Direct clinical validation is not part of the initial phase, but collaboration with a cardiologist is identified as a future step to assess medical accuracy and real-world applicability.
Why This System is Needed
Cardiac diagnosis often requires synthesizing information from multiple heterogeneous sources. ECG patterns, ventricular function measurements, laboratory values, medications, and medical history must be interpreted together to support safe and accurate decisions.
While experienced clinicians are trained to perform this integration manually, a structured AI-based system can help organize and contextualize data, reduce cognitive load, and improve consistency in reporting.
The system will support automated ECG analysis for arrhythmia detection and clinically relevant abnormalities, structured extraction and quantification of echocardiographic parameters, and contextual integration of medical history, medications, and laboratory results.
The objective is not to provide autonomous diagnosis, but to demonstrate how multimodal AI systems can assist clinicians by generating transparent, structured, and review-ready reports.
How We Plan to Achieve It
In order to achieve the goal, we divide the project objectives in 4 different sections:
1. Technical Requirements and System Definition
The first phase will define the regulatory scope of the system based on MDR/IVDR requirements and internal corporate templates. The core document types (e.g., requirements specifications, risk management files, verification reports) will be analyzed to identify mandatory fields, relationships, and traceability expectations. A data model will be defined to represent requirements, hazards, specifications, test cases, and their relationships as a structured traceability graph. Validation rules for uniqueness, cross-referencing, and coverage will be specified at this stage.
2. Architecture and Pipeline Design
The system architecture will be designed to support secure and efficient multimodal processing. Separate preprocessing pipelines will be developed for ECG signals, echocardiographic reports, and EHR data (via FHIR-compatible formats). These pipelines will normalize and structure heterogeneous inputs before inference.
Particular emphasis will be placed on efficient multimodal inference, deterministic output generation, and explainability. The inference layer based on MedGemma will be orchestrated to produce structured diagnostic summaries grounded explicitly in source data. Secure in-memory handling will be enforced across all components to guarantee compliance with privacy constraints. A standardized diagnostic report template will be defined to ensure consistency and interpretability.
3. Prototype Implementation
During this phase, the ECG analysis module, echocardiographic quantification module, and EHR integration layer will be implemented and integrated into a unified offline prototype.
The system will generate structured and explainable clinical summaries through an end-to-end workflow. A lightweight user interface will be developed to visualize ECG annotations, extracted echo parameters, and contextual patient information, and to generate structured diagnostic reports ready for clinical review. The outcome of this phase will be a working offline prototype demonstrating technical feasibility and robustness in controlled evaluation settings.
4. Testing, Validation, and Documentation
The prototype will be evaluated using public cardiac datasets under controlled conditions. Performance will be measured in terms of detection accuracy, processing time, stability, and consistency of generated reports. Comprehensive technical documentation will describe the architecture, preprocessing strategies, inference workflow, privacy mechanisms, and evaluation results. The project will conclude with a validated engineering prototype, structured diagnostic reporting capability, and a documented roadmap for future development.
Future work will include collaboration with a cardiologist to validate diagnostic outputs, refine report usability, and assess clinical reliability in real-world scenarios.
Project Timeline
- Technical Requirements and System Definition: 40-50 hours
- Architecture and Pipeline Design: 80–90 hours
- Prototype Implementation: 100–120 hours
- Testing, Validation, and Documentation: 50-60 hours
Total Time Frame: 270-320 hours