Machine learning framework for early detection of polio outbreaks from acute flaccid paralysis surveillance data.

Poliomyelitis remains a global public health concern despite remarkable progress toward eradication. The continued circulation of vaccine-derived polioviruses underscores the need for enhanced surveillance systems capable of early detection and rapid response. Acute Flaccid Paralysis (AFP) surveillance, a key component of the polio eradication, plays a vital role in detecting poliovirus transmission but is often limited by delays in data collection, analytical capacity, and fragmented reporting structures. Recent advances in artificial intelligence (AI) offer opportunities to address these challenges by improving data analysis, outbreak prediction, and decision-making. This study developed an AI-based predictive framework that utilizes AFP surveillance data to improve early poliovirus case detection. The dataset used to train the AI models integrates geographical, vaccination, clinical, and laboratory variables to capture the multifactorial determinants of poliovirus transmission. Ten machine learning algorithms representing tree-based, probabilistic, and neural approaches, along with their ensembles, were developed and compared. Among them, the CatBoost-based model achieved the highest performance, with an accuracy of 92.22% and an area under the operating curve of 0.99, surpassing both standalone and ensemble models. Model interpretability analyses using Shapley Additive Explanations and Local Interpretable Model-Agnostic Explanations identified body temperature, fatigue, and sore throat as the most influential predictors in suspected cases with AFP, consistent with early clinical indicators of poliovirus infection. The proposed machine learning framework demonstrates the value of integrating explainable AI into routine AFP surveillance to enhance outbreak prediction, support timely interventions, and strengthen global efforts toward the eradication of poliomyelitis.