buumba641/Kenya-Clinical-Reasoning

# Kenya Clinical Reasoning Challenge – Low‑Resource NLP for Healthcare This repository contains my solution for the **Kenya Clinical Reasoning Challenge** on Zindi. The goal is to predict a clinician’s free‑text response given a clinical vignette (nurse background + patient presentation). The evaluation metric is **ROUGE score**, and solutions must respect strict deployment constraints (≤1B parameters, <100ms inference, <2GB RAM). **Final scores:** - Public leaderboard: **0.35075** - Private leaderboard: **0.35945** - Benchmark (random/naive): 0.00271 / 0.00362 --- ## 📖 Problem Overview Frontline healthcare workers in rural Kenya face complex decisions with limited specialist support. The dataset consists of ~400 training and 100 test examples – authentic clinical prompts paired with expert clinician responses. The challenge simulates real‑world constraints: fast, accurate, and deployable on edge devices (e.g., NVIDIA Jetson Nano). **Key constraints:** - Model parameters ≤ 1 billion - Inference time < 100 ms per vignette - Inference RAM < 2 GB - Quantization required - Training ≤ 24 hours on T4 GPU --- ## 🧠 Approach ### 1. Model Architecture - **Base model:** `t5-base` (222M parameters) – well under the 1B limit. - Fine‑tuned as a sequence‑to‑sequence task: Input: `"summarize: " + enhanced_prompt` Output: clinician’s summary/assessment. ### 2. Preprocessing & Prompt Engineering - Cleaned text (standardised county names, removed problematic IDs). - Created an **enhanced prompt** by injecting metadata (nursing competency, clinical panel, years of experience) and patient age/gender when available. - All inputs prefixed with `"summarize: "` to align with T5’s pre‑training. ### 3. Training Strategy - **Stratified split** (80/20) based on clinician response length to preserve diversity. - Gradient accumulation (batch size 4, accumulation steps 8) to fit in T4 memory. - FP16 mixed precision. - Optimizer: AdamW (lr=3e-4) with cosine warmup. - Early stopping b …