Interpretable epidemic state estimation via rule based modeling

Epidemiological data is, by its very own nature, associated with noise and incompleteness, posing a significant challenge to extracting meaningful insights. To address this limitation, we propose a novel framework that seamlessly integrates fuzzy logic and machine learning techniques to provide a reliable understanding of the aforementioned data. Fuzzy logic, with its inherent ability to handle vagueness and imprecision, proves invaluable in interpreting noisy epidemiological data.

Our approach introduces a novel perspective by departing from a dynamical epidemic model, which encodes comprehensive prior knowledge of epidemic dynamics, to create a credible context for our predictive task. This facilitates model's outputs interpretability, while maintaining its credibility. Within this environment, we aim to detect relevant parts of the system state that are implicitly encoded in other state variables, by looking at observable variables of the epidemic state.

The Takagi-Sugeno model demonstrated robust predictive accuracy across varying Signal-Noise Ratio levels, achieving comparable performance to neural networks while maintaining interpretability, with significant advantages in noisy data scenarios, as evidenced by lower Root Mean Squared Errors under worst-case conditions (low SNRs).

This study introduces a robust and interpretable hybrid framework for epidemic forecasting, demonstrating reliable estimation of the effective reproduction number through fuzzy clustering and Takagi-Sugeno modeling, even under noisy conditions. The method effectively addresses data uncertainty and demonstrates strong performance under noisy conditions, offering a promising approach for applications requiring both transparency and reliability in predictive modeling.