B-Omare/zeramatumizi

A Causal AI Early Warning and Precision Intervention System for Drug and Substance Use Disorders in Kenya # ZeraMatumizi 🇰🇪 > *Zera Matumizi — "Eliminate Use" in Swahili* A longitudinal, open-science, production-grade **Causal AI Early Warning and Precision Intervention System** for Drug and Substance Use Disorders in Kenya — combining causal inference, Bayesian modelling, NLP, graph neural networks, and quantum computing into a unified public health intelligence platform. --- ## The Problem Over **1.5 million Kenyan youths** are grappling with drug and substance abuse. Over 90% of rehabilitation facilities are privately owned, skewed toward urban centres, and unaffordable to the majority of Kenyans. **No data-driven early identification system exists at the county level.** ZeraMatumizi addresses this gap by: 1. **Predicting** which individuals and communities are at highest risk — before clinical presentation 2. **Explaining** the causal pathways driving risk across 47 counties 3. **Optimising** allocation of Kenya's scarce treatment resources 4. **Generating** actionable intelligence for NACADA officers in both Swahili and English --- ## System Architecture Raw Data (KDHS 2022, NACADA, DHIS2, OSM) │ ▼ D1: ETL Pipeline ────────────────────────────────────────────── │ loader.py → validator.py → cleaner.py │ 4,000 respondents, 13 features, Pandera schema validation ▼ D2: Causal Inference ────────────────────────────────────────── │ dag.py → Interactive causal DAG (26 nodes) │ did_analysis.py → NACADA campaigns: -27% disorder rate │ rdd_analysis.py → Age-18 threshold: +56% disorder risk │ iv_analysis.py → Chang'aa proximity IV: β=0.557 ▼ D3: Bayesian Hierarchical Model ─────────────────────────────── │ hierarchical_model.py → County risk with credible intervals │ Nyamira 14.1% [9.7%, 20.7%] ... Homa Bay 11.7% [7.5%, 16.9%] ▼ D4: NLP & LLM Pipeline ──────────────────────────────────────── │ swahili_ner_model.py → Swahili SUD NER (5 entity types) │ rag_pipeline.py → NACADA counsellor RAG assistant │ topic_modelling.py → BERTopi …