Dynamic temporal partitioning enhanced transformer for pediatric viral load forecasting

Introduction Viral infectious diseases are highly prevalent in pediatric populations, and accurate prediction of viral load is critical for clinical intervention and public health management. Existing time-series models poorly handle long-range dependencies, multi-scale features, noise, and incomplete clinical data. Methods We propose DTR-Former, a Dynamic Temporal Partitioning-enhanced Transformer. It uses wavelet packet decomposition for adaptive temporal partitioning and multi-scale feature extraction, adopts sparse self-attention to reduce redundancy and enhance long-sequence modeling, and employs a residual convolutional decoder with a gating mechanism to refine features and suppress residual noise. Results On dbEBV dataset, DTR-Former achieves MSE = 0.16, MAE = 0.27, R 2 = 0.88; on NCBI dataset, MSE = 0.18, MAE = 0.30, R 2 = 0.86. It shows strong multi-step stability and robustness under missing-data conditions. Conclusion DTR-Former outperforms state-of-the-art methods in accuracy, stability, and efficiency, offering an effective solution for pediatric viral load forecasting and related time-series tasks.