Accurate clinical prediction using Electronic Health Records (EHRs) is crucial for improving patient outcomes. Current EHR representation approaches include sequential models that capture temporal dependencies and graph-based methods that model structural relationships, but both have limitations. Sequential models overlook complex clinical event relationships, while graph-based approaches inadequately incorporate temporal information. Moreover, existing methods primarily use discrete medical codes, neglecting continuous data like vital signs and lab results. We propose Personalized Multivariate Temporal Graphs (PMTG), a novel framework that encodes patient EHRs as individualized temporal graphs. Unlike traditional knowledge-graph approaches, PMTG integrates comprehensive clinical features-including lab values, vital signs, and assessment scores-alongside standard coded events. Our multigranular temporal encoding transforms timestamped events into time-aware features, capturing complex temporal dependencies across different granularities. We introduce the Dynamic Temporal Graph Transformer (DTGT) to model variable-sized dynamic graphs, accommodating diverse real-world patient records. This architecture extends the Transformer paradigm to jointly learn dynamic graph structure evolution and extract multi-granular temporal features from personalized graphs. Evaluated on MIMIC-IV for mortality and ICU readmission prediction, PMTG achieves state-of-the-art performance with AUROC of$\mathbf{8 0. 4 \%}$for mortality prediction and 90.7 % for ICU readmission prediction, significantly outperforming baselines. These results demonstrate PMTG's effectiveness in leveraging full-spectrum longitudinal EHR data for superior clinical prediction.

In skeleton-based gesture recognition tasks, existing approaches based on graph convolutional networks (GCNs) struggle to capture the synergistic actions of nonadjacent graph nodes and the information conveyed by their long-range dependencies. Combining spatial and temporal transformers is a promising solution to address the limitation, inspired by the advantage of transformer in assessing nonadjacent long-range dependencies, but there lacks an effective strategy to integrate the spatial and temporal information extracted by these transformers. Therefore, this article proposes the spatial-temporal alternating graph convolution transformer (ST-GCN-AltFormer), which connects the spatial-temporal graph convolutional network (ST-GCN) with the spatial-temporal alternating transformer (AltFormer) architecture. In the AltFormer architecture, the spatial-temporal transformer branch employs a spatial transformer to capture information from specific frames, and uses a temporal transformer to analyze its evolution over the entire temporal range. Meanwhile, the temporal-spatial transformer branch extracts temporal information from specific nodes using a temporal transformer, and integrates it with a spatial transformer. The fusion enhances accurate spatial-temporal information extraction. Our method achieves superior performance compared to state-of-the-art methods, achieving accuracies of 97.5%, 95.8%, 94.3%, 92.8%, and 98.31% on the large-scale 3D hand gesture recognition (SHREC’17 Track), Dynamic Hand Gesture 14-28 (DHG-14/28), and leap motion dynamic hand gesture (LMDHG) dynamic gesture datasets, respectively.

Molecular docking, a technique for predicting ligand binding poses, is crucial in structure-based drug design for understanding protein-ligand interactions. Recent advancements in docking methods, particularly those leveraging geometric deep learning (GDL), have demonstrated significant efficiency and accuracy advantages over traditional sampling methods. Despite these advancements, current methods are often tailored for specific docking settings, and limitations such as the neglect of protein side-chain structures, difficulties in handling large binding pockets, and challenges in predicting physically valid structures exist. To accommodate various docking settings and achieve accurate, efficient, and physically reliable docking, we propose a novel two-stage docking framework, DeltaDock, consisting of pocket prediction and site-specific docking. We innovatively reframe the pocket prediction task as a pocket-ligand alignment problem rather than direct prediction in the first stage. Then we follow a bi-level coarse-to-fine iterative refinement process to perform site-specific docking. Comprehensive experiments demonstrate the superior performance of DeltaDock. Notably, in the blind docking setting, DeltaDock achieves a 31\% relative improvement over the docking success rate compared with the previous state-of-the-art GDL model DiffDock. With the consideration of physical validity, this improvement increases to about 300\%.