Infectious disease outbreaks in crop systems pose a significant threat to global food security, particularly when early detection and intervention opportunities are missed. Predicting not just whether but when an infection will emerge is critical for effective disease management and control. In this context, survival analysis—a statistical framework for modeling time-to-event data—offers a natural and powerful solution. Widely adopted in epidemiology and clinical research, survival analysis can be adapted to plant disease surveillance for predicting the timing of infection onset at the host or population level. However, deep survival models based on multilayer perceptrons (MLPs) often struggle with high-dimensional agricultural data, leading to issues such as overfitting, poor parameter efficiency, and limited interpretability. To address these challenges, we introduce a Kolmogorov–Arnold Network (KAN) architecture into the survival analysis context, leveraging its compact nonlinear function-approximation capabilities to improve predictive performance and efficiency. We present Rek-Surv, a lightweight deep survival model built on an Efficient-KAN backbone and augmented with residual connections and enhanced regularization. Rek-Surv is evaluated on five clinical benchmark datasets as well as a citrus Huanglongbing (HLB) plant disease dataset, demonstrating its generalizability across human and plant infectious disease contexts. On the HLB task, Rek-Surv achieves a high concordance index (C-index) of 0.962 with only 114 trainable parameters and millisecond-level inference speed. It outperforms existing survival models by providing more accurate outbreak onset predictions with a fraction of the model complexity. This efficiency makes Rek-Surv well-suited for real-time outbreak detection and early intervention, illustrating how advanced survival models can enable proactive infectious disease management in both agriculture and healthcare.
Xiao et al. (Sun,) studied this question.