Leveraging large language models for gastrointestinal injury detection in athletes: a medical image analysis approach

Figure 1. Overview of the BANNet architecture. This figure illustrates the overall structure of the proposed Biomechanical-Aware Neural Network (BANNet). Three multimodal data streams–biomechanical signals, medical imaging, and contextual features–are independently encoded and fed into the hierarchical BANNet pipeline. The left module performs spatiotemporal fusion to integrate sequential and spatial cues. The middle module applies cross-attention and self-attention mechanisms to refine multimodal feature representations. The right module conducts injury risk estimation using calibrated prediction heads and uncertainty-aware inference. The architecture enables robust multimodal understanding and reliable injury forecasting in athlete health monitoring.

Figure 2. Multimodal feature encoding in BANNet. This figure demonstrates how BANNet processes heterogeneous inputs through modality-specific encoders. Biomechanical signals are modeled using a GRU encoder to capture temporal dependencies. Medical imaging data are extracted using convolutional neural networks to obtain spatial representations. Contextual variables, including training load and environmental conditions, are embedded using a fully connected network. The encoded features are aligned and integrated within the Multimodal Fusion Encoder, forming unified latent embeddings that support downstream spatiotemporal fusion and injury risk estimation.

Figure 3. Overview of the ARPOS framework for adaptive rehabilitation and performance optimization. The Adaptive Rehabilitation and Performance Optimization Strategy (ARPOS) integrates sensor-based measurements, reinforcement learning, and graph-based coordination to deliver personalized rehabilitation protocols. The left module extracts informative biomechanical and physiological features via a Pyramid Reduction and Multi-headed Self-Attention encoder. The central module employs reinforcement learning to generate adaptive interventions based on real-time feedback. The rightmost module introduces graph-based coordination, enabling synchronized optimization across multiple athletes in team-based settings. ARPOS ensures dynamic, individualized, and context-aware rehabilitation planning..

Figure 4. Reinforcement learning-based intervention optimization in ARPOS. This figure details the reinforcement learning subsystem responsible for optimizing rehabilitation interventions. Biomechanical and contextual data are first processed to form structured feature inputs. The central RL module, based on proximal policy optimization (PPO) and heuristic refinement, learns adaptive intervention strategies to enhance recovery quality. The right section illustrates the scatter–gather refinement mechanism, which ensures stable and continuous optimization of rehabilitation policies. This RL-driven module forms the core of ARPOS’s ability to adjust interventions dynamically in response to real-time sensor feedback.