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

Abstract

Humanitarian settings are highly vulnerable to infectious disease outbreaks because displacement, crowding, disruption of health services, insecurity, and inadequate water and sanitation often interact in ways that are difficult to capture with conventional prediction models. There is a need for forecasting approaches that can integrate heterogeneous data sources and better represent complex system structure.

We developed and evaluated a machine-learning framework incorporating topological data analysis to predict cholera and measles surge events (binary indicators per LGA-week) across 97 Local Government Areas in Nigeria between 2018 and 2023. These 97 LGAs represent a selected high-burden subset (12.5%) of Nigeria’s 774 LGAs with sufficient surveillance data. Weekly district-level predictors included climate, conflict, displacement, health-system, and socioeconomic variables. Persistent homology was used to derive topological summaries from multivariate risk profiles, and these were combined with selected raw predictors in gradient-boosting models. Outcomes were defined using surveillance-based outbreak thresholds with a 4-week prediction horizon. Model performance was assessed using temporally ordered hold-out validation, with evaluation of discrimination, calibration, and incremental value over baseline models.

The topological models achieved ROC-AUC of 0.78 (95% CI: 0.74–0.82) for cholera and 0.81 (95% CI: 0.77–0.85) for measles, representing modest improvements of 0.08–0.12 over models using only conventional predictors. At the optimal decision threshold determined using Youden’s index on validation data, sensitivity was 0.72 (range across folds: 0.68–0.76) and specificity was 0.82 (range: 0.79–0.85) for cholera, with false alert rates varying from 2.8–3.6 per LGA per year across temporal folds. Topological features contributed 35% of predictive importance. Calibration slopes were 0.94 (cholera) and 0.97 (measles).

Topological feature representations provide a modest but meaningful complementary approach for outbreak prediction in complex humanitarian environments. Their value appears to lie in summarizing higher-order structure across multiple interacting risk domains, rather than replacing established epidemiologic indicators. However, routine deployment requires prospective validation and context-specific threshold tuning. Further external validation, operational threshold analysis, and prospective testing are needed before routine deployment in public-health early warning systems.

1. Introduction

Infectious disease outbreaks remain a recurrent and often preventable source of morbidity and mortality in humanitarian settings. Forced displacement, conflict, environmental shocks, and health-system disruption create conditions that increase exposure risk while simultaneously weakening detection and response capacity. According to UNHCR, 123.2 million people were forcibly displaced worldwide by the end of 2024, underscoring the scale at which outbreak-prone humanitarian conditions now intersect with global health security [1]. In such contexts, overcrowding, population mobility, interrupted vaccination, and poor access to safe water and sanitation can amplify transmission of epidemic-prone diseases and delay containment.