TDA-PIDO: A topological data analysis approach for early warning of infectious disease outbreaks

Public summary

* Our method maps disease time series to topological spaces using TDA with a noise-robust framework.

* Our method validates on COVID-19/SARS data, achieving robust outbreak classification beyond state of the art.

* Our method maintains high accuracy with sparse data, showing noise-robust superiority.

* Our method uses classifier-agnostic end-to-end deep learning, removing predefined model dependency.

Abstract

Infectious disease outbreak prediction is critical for mitigating healthcare systemic risks. However, noisy surveillance data severely degrade the prediction accuracy and limit model generalizability. Owing to the high robustness of topological data analysis (TDA) to noise, this study overcomes this challenge by introducing a TDA-based feature representation approach, TDA-PIDO (TDA for the prediction of infectious disease outbreaks). TDA-PIDO first converts different time series into respective point clouds. It then systematically extracts topological features from the point clouds. To enhance generalization, models are trained on simulated data generated from two noise-induced susceptible-infected-recovered (SIR) models and subsequently applied to real-world infectious disease datasets. Experimental results show that TDA-PIDO outperforms existing methods on simulated datasets with the same feature dimensionality and achieves greater performance on real-world data. Moreover, experiments at varying noise levels demonstrate the framework’s strong robustness. Overall, this work highlights TDA as a promising analytical tool for early warning.