Multimodal Heart Disease Diagnostic Pipeline

An advanced, end-to-end clinical decision-support system that integrates multimodal AI models to evaluate heart disease risk. This pipeline combines visual analysis of ECGs, tabular clinical data processing, and large-scale medical reasoning to generate explainable diagnostic reports.

🚀 Overview

The pipeline employs a "Specialized-to-Generalist" architecture. It uses upstream domain-specific models to extract high-level features from diverse data sources, which are then synthesized by a clinical LLM to produce a final medical evaluation.

Key Components

• Visual Analysis: GemmaECG-Vision for 12-lead ECG waveform interpretation.
• Tabular Risk: Google TabFM for metabolic and clinical metric risk scoring.
• Clinical Embeddings: Bio_ClinicalBERT for medical text representation.
• Reasoning Engine: BioMistral-7B for multimodal synthesis and report generation.

✨ Features

• Multimodal Integration: Seamlessly combines image (ECG) and tabular (Clinical) data.
• Explainable AI: Generates a dual-format report (one for clinicians, one for patients).
• Diagnostic Concurrence: Explicitly validates the agreement between statistical risk (TabFM) and holistic clinical reasoning (BioMistral).
• Production-Ready: Full Docker support with memory-optimized loading (float16) and lazy loading for efficient deployment.

🛠️ Installation & Deployment

Prerequisites

• Docker
• NVIDIA GPU (recommended for inference speed)

Local Deployment

1. Clone the repository:

   git clone https://github.com/[your-username]/Heart_Disease.git
   cd Heart_Disease
   

2. Build the Docker image:

   docker build -t heart-disease-diag .
   

3. Run the container:

   docker run -p 7860:7860 heart-disease-diag
   

Hugging Face Spaces Deployment

The /Heart directory is optimized for Hugging Face Spaces. To deploy:

1. Upload the contents of the Heart/ folder to a new Space.
2. Select the Docker SDK.

📖 Documentation

For a detailed explanation of the architecture, the 23 input metrics, and the endpoint specifications, please refer to the System Overview.

📡 API Usage

Endpoint: POST /predict

Input:

• 12-lead ECG Image (File)
• 21 Patient Metrics (JSON)

Output:

• Comprehensive Clinical Report
• Patient-Friendly Summary
• Upstream model findings (ECG descriptions & TabFM risk %)

---

Disclaimer: This system is designed for educational and research purposes and should be used as a decision-support tool, not as a replacement for professional medical diagnosis.