TempAD-LesionNet: Adaptive Temporal Attention for Cranial Lesion Forecasting in Neurocritical Care via Seq2Seq Architectures

Cranial lesion progression presents significant clinical risks in neurocritical care. However, early prediction remains challenging due to the irregularity and heterogeneity of ICU data. In this paper, we propose TempAD-LesionNet, a novel sequence-to-sequence (Seq2Seq) model that incorporates a temporal attention mechanism and an adaptive hybrid decoder. The decoder dynamically selects between LSTM and Transformer layers based on the input variance. TempAD-LesionNet is trained using a multi-objective loss function that jointly optimizes lesion volume, shape, and intensity. We evaluate the model on three real-world datasets (MIMIC-IV, eICU, and ICP-ICU). It achieves a Mean Absolute Error (MAE) of 0.095, a Concordance Index (C-Index) of 0.82, and a Dynamic Time Warping (DTW) score of 1.94—outperforming baseline models such as RNN, Informer, and TimesNet. Temporal attention maps confirm the interpretability of the model’s focus, while survival analysis demonstrates its utility for clinical risk stratification. These results suggest that TempAD-LesionNet is well-suited for real-time application in neurocritical care environments.