In order to provide more information and explanation about our training data and evaluation benchmark, we introduce our synthetic samples with respect to the length and density of point trajectory.

Recent advances in transformer architectures have revolutionised natural language processing, but their application to healthcare domains presents unique challenges. Patient timelines are characterised by irregular sampling, variable temporal dependencies, and complex contextual relationships that differ substantially from traditional language tasks. This paper introduces \METHOD~(Modular Efficient Transformer for Health Outcome Discovery), a novel transformer architecture specifically designed to address the challenges of clinical sequence modelling in electronic health records. \METHOD~integrates three key innovations: (1) a patient-aware attention mechanism that prevents information leakage whilst enabling efficient batch processing; (2) an adaptive sliding window attention scheme that captures multi-scale temporal dependencies; and (3) a U-Net inspired architecture with dynamic skip connections for effective long sequence processing. Evaluations on the MIMIC-IV database demonstrate that \METHOD~consistently outperforms the state-of-the-art \ETHOS~model, particularly in predicting high-severity cases that require urgent clinical intervention. \METHOD~exhibits stable performance across varying inference lengths, a crucial feature for clinical deployment where patient histories vary significantly in length. Analysis of learned embeddings reveals that \METHOD~better preserves clinical hierarchies and relationships between medical concepts. These results suggest that \METHOD~represents a significant advancement in transformer architectures optimised for healthcare applications, providing more accurate and clinically relevant predictions whilst maintaining computational efficiency.

Enabling LLMs to handle lengthy context is currently a research hotspot. Most LLMs are built upon rotary position embedding (RoPE), a popular position encoding method. Therefore, a prominent path is to extrapolate the RoPE trained on comparably short texts to far longer texts. A heavy bunch of efforts have been dedicated to boosting the extrapolation via extending the formulations of the RoPE, however, few of them have attempted to showcase their inner workings comprehensively. In this paper, we are driven to offer a straightforward yet in-depth understanding of RoPE extensions from an attention perspective and on two benchmarking tasks. A broad array of experiments reveals several valuable findings: 1) Maintaining attention patterns to those at the pretrained length improves extrapolation; 2) Large attention uncertainty leads to retrieval errors; 3) Using longer continual pretraining lengths for RoPE extensions could reduce attention uncertainty and significantly enhance extrapolation.