KESOZI Digital Twin is an ongoing scientific machine learning project that develops a Physics-Informed Multimodal Artificial Intelligence framework for estimating, forecasting, and simulating childhood diarrheal disease burden across Africa. The platform integrates Physics-Informed Neural Networks (PINNs), Graph Neural Networks (GNNs), epidemiological diffusion-reaction modeling, multimodal data fusion, pathogen-attribution analytics, uncertainty quantification, and Digital Twin simulation to model disease transmission dynamics, spatial propagation, outbreak trajectories, and intervention impacts.
The framework combines epidemiological, climate, environmental, healthcare-access, nutritional, water-sanitation-hygiene (WASH), and pathogen-related datasets to create a virtual representation of disease behavior in Kenya, Somaliland, and Zimbabwe. By linking mechanistic epidemiological processes with advanced artificial intelligence, the project provides a physics-grounded and uncertainty-aware platform for infectious disease surveillance, climate-health analytics, and public-health decision support.
To develop a reproducible Digital Twin framework capable of independently estimating and predicting childhood diarrheal disease burden while preserving epidemiological consistency through physics-informed learning.
The framework aims to:
Estimate under-five morbidity and mortality burden.
Predict temporal outbreak dynamics and seasonal disease variation.
Model spatial transmission and geographic disease hotspots.
Quantify pathogen-attributed disease burden.
Evaluate climate-sensitive disease propagation.
Simulate intervention scenarios and public-health strategies.
Generate uncertainty-aware forecasts for evidence-based decision making.
Diffusion-reaction epidemiological equations describing disease transmission, recovery, and spatial propagation are embedded directly into neural-network optimization, enabling physically consistent forecasting even under sparse surveillance conditions.
PINNs learn disease trajectories while enforcing governing epidemiological constraints, reducing physically unrealistic predictions and improving generalization.
Graph-based learning models interactions among neighboring counties and regions, capturing spatial clustering, transmission corridors, and outbreak propagation pathways.
The framework integrates:
Epidemiological surveillance data
Climate indicators
Environmental risk variables
WASH indicators
Healthcare-access metrics
Nutritional risk indicators
Population statistics
Pathogen prevalence information
to produce comprehensive disease-risk estimates.
The system estimates mortality and morbidity contributions from major enteric pathogens including:
Rotavirus
Shigella
Cryptosporidium
Norovirus GII
ST-ETEC
Cholera
Campylobacter
Adenovirus 40/41
Graph-based neighborhood risk propagation identifies disease hotspots and geographically clustered transmission zones.
Prediction intervals and mortality uncertainty estimates provide confidence bounds for epidemiological forecasts and policy planning.
Virtual intervention scenarios evaluate:
Flood shocks
WASH improvements
Healthcare interventions
Combined public-health strategies
allowing assessment of intervention effectiveness before real-world implementation.
Higher composite risk scores were strongly associated with increased childhood diarrheal incidence, demonstrating the importance of integrating environmental, nutritional, healthcare, and WASH determinants.
Climate forcing exhibited substantial temporal variability across Kenya, Somaliland, and Zimbabwe, indicating strong environmental influences on disease transmission dynamics.
The framework estimated substantial differences in under-five diarrheal mortality burden across study settings, reflecting variations in sanitation, healthcare accessibility, environmental conditions, and population characteristics.
Somaliland exhibited the highest population-adjusted diarrheal incidence rates, highlighting significant environmental and healthcare vulnerabilities.
Rotavirus emerged as the dominant contributor to modeled mortality burden across all study countries, followed by Shigella and Cryptosporidium.
Graph-learning analysis identified geographically concentrated disease-risk clusters and potential transmission hotspots.
Physics-Informed Neural Networks successfully captured underlying disease dynamics while filtering surveillance noise and maintaining stable convergence during training.
Intervention simulations demonstrated that combined WASH and healthcare interventions produced the greatest reduction in projected childhood diarrheal mortality, outperforming individual intervention strategies.
Uncertainty analysis generated stable prediction intervals, supporting robust and reliable epidemiological forecasting.
Childhood diarrheal disease remains one of the leading causes of mortality among children under five in many African countries. Traditional epidemiological forecasting systems often struggle with sparse surveillance data, weak spatial representation, and limited mechanistic understanding.
KESOZI Digital Twin addresses these limitations by combining epidemiological physics, artificial intelligence, multimodal data fusion, graph learning, uncertainty quantification, and digital-twin simulation into a unified framework.
The platform enables:
Early outbreak detection
Climate-sensitive disease forecasting
Geographic hotspot identification
Pathogen burden estimation
Public-health intervention planning
Resource allocation optimization
Evidence-based policy development
for resource-constrained healthcare environments.
Physics-Informed Neural Networks (PINNs)
Graph Neural Networks (GNNs)
Ensemble Learning Models
Multimodal Fusion Architectures
Diffusion-Reaction Disease Models
Susceptible-Infected-Recovered (SIR) Dynamics
Disease Burden Estimation Models
Pathogen Attribution Frameworks
World Health Organization (WHO)
UNICEF
World Bank
Our World in Data
Climate and Environmental Datasets
WASH Indicators
Healthcare Accessibility Metrics
Python
NumPy
Pandas
TensorFlow / PyTorch
Network Analysis Libraries
Geospatial Analytics
Scientific Machine Learning Frameworks
Current Status: Proof-of-concept framework completed and manuscript prepared.
Integration of real-world surveillance datasets.
Calibration against country-level epidemiological records.
Real-time disease surveillance data assimilation.
Expansion to additional African countries.
Incorporation of satellite-derived environmental intelligence.
Mobile and web-based Digital Twin deployment.
Integration with public-health early warning systems.
Development of intervention optimization and decision-support modules.
Validation using prospective epidemiological monitoring data.
To establish a continental-scale AI-powered Digital Twin platform for infectious disease forecasting, climate-health analytics, outbreak preparedness, and public-health decision support across Africa