Grandkojo/feverdiff_ai

AI-powered differential diagnosis for malaria, dengue, chikungunya, and typhoid in Ghana's resource-limited settings using fine-tuned MedGemma 1.5 # 🏥 FeverDiff AI: Differential Diagnosis Assistant **FeverDiff AI** is a MedGemma-powered clinical decision support tool designed to differentiate between four common causes of fever in Ghana: **Malaria, Dengue, Chikungunya, and Typhoid**. --- ## 🚀 Overview In Ghana, overlapping symptoms lead to dangerous misdiagnosis. FeverDiff AI bridges the gap between rural clinics and specialist support using a hybrid AI approach. ### Key Features - **MedGemma 1.5 4B + Gemini 2.5-Flash**: Core clinical reasoning powered by specialized medical LLMs. - **Ghana-Specific Clinical Scoring**: Tailored logic for regional epidemiological patterns. - **High-Fidelity Dashboard**: Comprehensive data entry for 26+ clinical variables. - **Researcher Workspace**: Direct JSON import for rapid test scenario validation. - **GPU-Powered Backend**: Scalable serverless inference via Modal Labs. --- ## 📸 Screenshots | Dashboard Overview | Clinical Reasoning | Statistical Distribution | | :---: | :---: | :---: | | | | | --- ## 🛠️ Tech Stack - **Frontend**: Vanilla HTML5, CSS3 (Glassmorphism), JavaScript (ES6+). - **Backend API**: FastAPI on Modal Labs (GPU T4). - **AI Models**: MedGemma 1.5 4B (Quantized 8-bit), Gemini 2.5-flash for reasoning refinement. - **Data**: 10,000 synthetic clinical records validated by WHO/CDC guidelines. --- ## 📂 Project Structure - `modal_backend.py`: GPU serving logic with volume caching. - `model.py`: Diagnostic engine and prompt engineering. - `feverdiff_opt.ipynb`: Model optimization and clinical rule validation notebook. - `static/`: High-fidelity UI assets. - `test_scenarios.json`: Pre-defined clinical cases for validation. --- ## 🧠 Clinical Intelligence Engine FeverDiff AI employs a multi-layered diagnostic approach combining hard clinical rules with targeted LLM prompting. ### 1. Unified Targeted Prompting (UTP) The model receives a structured clinical profile rather than a conversational query. The prompt is engineered to: - **Enforce …