Open Source Project

Cardio-Sahayak India 🇮🇳 🫀

A Specialized Multimodal LLM for Complex Cardiology Care

Optimized for the South Asian demographic, addressing unique genetic predispositions, clinical profiles, and socio-economic contexts to democratize advanced cardiac care.

Try Live Demo View Repository Read Preprint

The South Asian Cardiac Challenge

Myocardial Infarction (MI) in India occurs 5-10 years earlier than in Western populations. Despite standard BMIs, central obesity, genetic predispositions (like the MYBPC3 Δ25bp variant), and unique metabolic profiles present a distinct "South Asian Phenotype" often overlooked by generalized medical AI models.

🧬 Population-Specific Needs

Current models fail to account for lower BMI thresholds for MI, elevated Lp(a) levels, and specific variants prevalent in up to 4% of the Indian population.

Dual-Architecture AI

Combining MedGemma-27B for deep clinical reasoning and a Vision-Language Model (MedSigLIP) for native 12-lead ECG interpretation.

📊 Two-Phase Contextualization

Fine-tuned sequentially on foundational ICMR guidelines (Phase 1) and then on a rigorously curated dataset of real Indian clinical notes and Gemini-synthesized phenotype shifts (Phase 2).

Technical Architecture

Cardio-Sahayak implements a state-of-the-art dual-architecture design:

  • Text & Reasoning Backbone: google/medgemma-27b-it, quantized using 4-bit NormalFloat (NF4) via bitsandbytes and fine-tuned using QLoRA.
  • Multimodal ECG Processing: google/medsiglip-448 integrated via the MEIT framework, allowing raw 12-lead ECG ingestion and cross-modal diagnostic reasoning.
  • Data Pipeline: The V2 dataset explicitly integrates ekacare clinical notes, synthetic South Asian phenotype shifts, and digitized ECG mocks to deeply embed Indian cardiac contexts.

Democratizing Expert Care (Offline-Ready)

Through a custom architectural patch bypassing the complex Gemma3 VLM structure, we successfully converted the fine-tuned model into a highly compressed Q4_K_M GGUF format (16.6GB). Cardio-Sahayak India is fully capable of running completely offline on standard clinical laptops in resource-constrained rural clinics.

Try Live Demo
Explore the Codebase on GitHub →

🔓 Open Source

Released to the community to foster collaboration, rigorous evaluation, and advancements in global health equity.

🌍 Targeted Impact

A specialized solution for the 1.4 billion people of India, directly addressing the region's rapidly growing burden of cardiovascular disease.