Topological data analysis for predicting disease outbreaks in humanitarian settings: A machine learning approach.

Area of Science:

• Epidemiology
• Machine Learning
• Topological Data Analysis

Background:

• Humanitarian settings face unique challenges in disease outbreak prediction due to factors like displacement, crowding, and disrupted health services.
• Conventional models struggle to capture the complex interactions of these risk factors.
• There is a need for advanced forecasting approaches that can integrate diverse data sources and represent complex system dynamics.

Purpose of the Study:

• To develop and evaluate a machine learning framework using topological data analysis for predicting cholera and measles surge events in Nigeria.
• To assess the performance and added value of topological features compared to conventional predictors.

Main Methods:

• A machine learning framework incorporating topological data analysis was developed.
• Persistent homology was used to derive topological summaries from multivariate risk profiles.
• Gradient-boosting models combined topological features with conventional predictors (climate, conflict, displacement, health-system, socioeconomic variables) to forecast surge events in Nigerian Local Government Areas (LGAs) with a 4-week horizon.

Main Results:

• Topological models achieved an ROC-AUC of 0.78 for cholera and 0.81 for measles, outperforming conventional models by 0.08-0.12.
• Sensitivity and specificity for cholera prediction were 0.72 and 0.82, respectively, at the optimal threshold.
• Topological features contributed 35% to predictive importance, indicating their significant role in forecasting.

Conclusions:

• Topological feature representations offer a valuable complementary approach for outbreak prediction in complex humanitarian environments.
• These methods effectively summarize higher-order structures across interacting risk domains.
• Further prospective validation and context-specific tuning are necessary for routine deployment in public health early warning systems.