Predicting Infectious Disease Risk with Graph Neural Networks and High-Accuracy Contact Data in a Mixed-Age Preschool

Background : Infectious disease outbreaks in schools, particularly in preschools, pose a significant public health challenge. While school closures can curb transmission, they are a broad and disruptive measure. Targeted interventions require a deeper understanding of fine-grained contact patterns and the associated transmission risk within these high-density settings. This study aims to construct high-resolution contact networks to identify high-risk populations, thereby enabling precise risk assessment and informing targeted control strategies. Methods : High-frequency (0.25 Hz) Ultra-Wideband (UWB) wearable devices were employed to accurately capture close-range (<2 meters) contact data among students and teachers in a mixed-age preschool. The data were used to construct undirected weighted contact networks. An improved graph convolutional network architecture (DeepGNNMinorityFocus) was then applied, using illness-related absenteeism days as the outcome, to develop a three-tier (low/medium/high) risk prediction model that quantifies transmission risk from individual to population levels. Results : The model identified high-risk features, primarily including third-grade students, male students, and individuals with prominent values in key topological metrics such as degree and closeness centrality. Network analysis further revealed that third-grade students were overrepresented in the top 10% of these key topological metrics, exhibiting the longest single-contact duration and highest contact frequency among all grades, yet their total contact duration and total number of contacts were the lowest. Contact patterns showed a high degree of in-class aggregation, with in-class contact duration accounting for 89.77% of the total contact time. In terms of teacher-student interactions, the total contact duration decreased significantly as grade level increased. The constructed risk prediction model demonstrated excellent performance in individual risk assessment, achieving an accuracy of 87.5% and a weighted F1-score of 0.875. Conclusions : The accurate identification of high-risk individuals in preschools provides a critical quantitative basis for designing targeted interventions and making informed adjustments to teaching models.