mrfayntom/kenya-clinical-reasoning

This project builds a lightweight AI model to generate clinician-like responses to medical case scenarios in Kenya. It supports healthcare workers in low-resource areas by providing accurate, structured, and context-aware clinical guidance using a fine-tuned Flan-T5 model. # AI Clinician Response Prediction – HealthTech Hackathon 2025 ## Overview This repository was developed for the HealthTech Hackathon 2025. The project focuses on building a machine learning model to replicate human clinician responses to clinical case scenarios (vignettes) commonly encountered in Kenyan healthcare settings. The aim is to support frontline decision-making in low-resource environments, where access to specialists and diagnostic tools is often limited. By modeling expert clinician responses, the system is intended to assist healthcare workers with accurate, reliable, and context-sensitive guidance. ## Dataset Description The dataset contains: - **Prompt**: A clinical vignette that presents a realistic medical case scenario. - **Response**: A written response by a qualified human clinician. Although the full dataset includes responses from large language models (LLMs), this challenge is focused solely on replicating the responses written by human clinicians. ## Objective The primary goal of this project is to: - Predict accurate and contextually appropriate clinician responses. - Simulate real-world medical decision-making in low-resource settings. - Enhance clinical support tools with human-aligned AI models. ## Suggested Approaches Possible modeling strategies include: - Fine-tuning pretrained language models (e.g., BERT, RoBERTa, GPT-style models). - Retrieval-augmented generation using medical knowledge sources. - Encoder-decoder architectures for response generation. - Ranking or scoring systems for response selection. ## Technology Stack - **Programming Language**: Python 3.10+ - **Core Libraries**: PyTorch, Hugging Face Transformers, scikit-learn, pandas, numpy - **Experiment Tracking**: Weights & Biases or TensorBoard - **Development Tools**: Jupyter Notebooks, Google Colab (optional) # Install dependencies ```bash pip install -q datasets accelerate bitsandbytes peft transformers ``` ## Model and Training The training pipeline …