AFTS: A patient-agnostic encoder-decoder architecture with directional attention for blood glucose forecasting

Accurate blood glucose forecasting remains challenging due to inter-patient heterogeneity and complex glycemic dynamics. We present AFTS (Adaptive Feature Time Series), a patient-agnostic deep learning architecture combining a bidirectional LSTM encoder-decoder with cascaded Directional Representation (DR) modules. These modules introduce a specialized axis-wise attention mechanism that processes temporal and feature dimensions separately, designed to disentangle trend evolution from latent feature magnitude. We evaluated AFTS on two real-world CGM datasets (KDD18 and CDD23) against twenty baseline models, including advanced Transformers and RNN variants. Under a rigorous patient-wise 80/20 split, AFTS achieved an MAE of 7.02 mg/dL (KDD18) and 7.39 mg/dL (CDD23) at a 30-minute prediction horizon. The results demonstrate that AFTS is numerically competitive with state-of-the-art architectures while offering a distinct mechanism for hierarchical feature refinement. By isolating the encoder-decoder backbone and DR modules in ablation studies, we confirm that the axis-wise attention mechanism contributes specifically to minimizing prediction error in complex glycemic scenarios. These findings establish AFTS as a robust architectural candidate for patient-agnostic forecasting, effectively balancing the capture of short-term fluctuations and long-term trends.