Neural Network Model Predicts Alternating Recurrent Events Like Mood Episodes

A new paper from Abigail Loe, Susan Murry, and Zhenke Wu introduces a neural network-based framework for dynamically predicting alternating recurrent events—event sequences where one occurrence triggers a secondary refractory period. These patterns appear in fields like behavioral science (e.g., mood episodes), criminal justice (e.g., recidivism cycles), and biostatistics (e.g., disease flare-ups). The challenge lies in correlated observations and repeated outcomes subject to potential censoring, which the authors address via inverse probability weighted pseudo-observations. This approach allows the model to account for time-varying factors while maintaining statistical rigor suitable for a broad audience.

The proposed method was validated through simulations showing good performance, then applied to a real-world dataset of first-year medical residents to predict periods of low mood. Results demonstrated outstanding predictive capability, suggesting the model can effectively forecast alternating events even with complex dependencies and censoring. The framework combines neural network flexibility with established survival analysis techniques, offering a practical tool for researchers and practitioners. By making the theory accessible and providing an online dynamic prediction approach, the authors enable real-time updates as new data arrives—a key requirement for clinical or justice system decision support.