Lafifi-24/-Kenya-Clinical-Reasoning-Challenge

# Kenya Clinical Reasoning Challenge Predicting clinical responses from Kenyan healthcare workers in resource-limited settings (T4 GPU) using machine learning. ## Overview This project addresses the Kenya Clinical Reasoning Challenge by training ML models to predict nurse responses to medical cases. The dataset contains 400 authentic clinical prompts—each combining a nurse's background with a complex medical situation from rural Kenyan healthcare facilities. The goal is to replicate the clinical reasoning of trained professionals working under pressure with limited resources and specialist support. ## Dataset - **Training samples**: 400 clinical vignettes - **Test samples**: 100 clinical vignettes - **Features**: Patient presentation, nurse experience level, facility type, medical domain - **Target**: Clinician's response to the scenario - **Domains**: Maternal and child health, critical care, infectious diseases, and more Each vignette reflects real-world constraints faced by healthcare workers in underserved regions of Kenya. ## Models We evaluate multiple architectures to find the best approach for this small, high-quality dataset: - **LSTM**: Sequence-based neural network for text generation - **T5**: Text-to-text transformer (t5-small, t5-base) - **Qwen3**: Alibaba's multilingual language model - **Llama 3.1**: Meta's open-source LLM - **Gemma**: Google's lightweight language model ## Evaluation Metrics - **ROUGE Score**: Primary metric for text generation quality - ROUGE-1: Unigram overlap - ROUGE-2: Bigram overlap - ROUGE-L: Longest common subsequence ## Results | Model | ROUGE-1 | ROUGE-2 | ROUGE-L | Notes | |-------|---------|---------|---------|-------| | T5 | TBD | TBD | TBD | Baseline | | LSTM | TBD | TBD | TBD | - | | Qwen3 | TBD | TBD | TBD | - | | Llama 3.1 | TBD | TBD | TBD | - | | Gemma | TBD | TBD | TBD | - |