prren-visual-risk-detection

# PRREN: Visual Recognition Framework for Operating Room Nursing Risk Factors This repository contains the implementation of **PRREN (Probabilistic Relational Risk Encoding Network)** — a computer vision-based framework for **real-time detection and mitigation of nursing risk factors in operating rooms (ORs)**. PRREN integrates convolutional and transformer-based architectures to detect safety-related visual events such as instrument mishandling, sterility breaches, and ergonomic risks. --- ## Overview Operating rooms are complex, safety-critical environments where nursing-related risks can compromise patient outcomes. **PRREN** provides an automated real-time risk assessment system that: - Detects and quantifies visual cues associated with unsafe nursing practices. - Models multi-agent interactions and contextual dependencies using relational attention. - Generates real-time risk alerts and supports adaptive interventions through the **Contextual Risk Intervention Mechanism (CRIM)**. --- ## Core Components | Module | Description ||--------|--------------|| **Relational Attention Mechanism (RAM)** | Models directional influence between nursing agents in multi-agent surgical scenes. || **Context-Aware Risk Encoding (CARE)** | Integrates environmental information such as workflow phase, patient stability, and workload. || **Temporal Smoothness Regularization (TSR)** | Stabilizes predictions and reduces noise in real-time video analysis. || **Contextual Risk Intervention Mechanism (CRIM)** | Converts continuous probabilistic risk signals into actionable alerts. | Refer to Figures 1–4 (pages 5–10) in the paper for detailed architecture diagrams. --- ## Methodology The PRREN framework combines:- **Convolutional Neural Networks (CNNs)** for spatial feature extraction. - **Transformer modules** for temporal and contextual modeling. - **Probabilistic relational inference** to predict risk propagation between agents. Training is conducted on multi-modal surgical datasets (video, sensor, and audio data), using evaluation metrics such as:- Accuracy, Recall, F1-Score - AUC-ROC - Mean Intersection-over-Union (mIoU) for workflow segmentation --- ## Datasets 1. **Operating Room Activity Dataset** – annotated surgical videos for activity recognition. 2. **Surgical Nursing Risk Factors Dataset** – multi-label risk dataset with contextual metadata. 3. **Visual Recognition for Medical Environments (VRME)** – large-scale visual dataset emphasizing occlusion and lighting variation. 4. **Operating Room Workflow Analysis Dataset** – synchronized video and sensor data for phase recognition and workflow modeling. (See Section 4.1 of the paper for detailed dataset descriptions.) --- ## Implementation Details - Framework: **PyTorch**- Hardware: **NVIDIA A100 (80 GB)**- Key parameters: - Learning rate: `5e-5` - Weight decay: `0.01` - Batch size: `8` - Sequence length: `64 frames` - Optimizer: `AdamW` with cosine annealing- Multi-modal fusion with early concatenation and cross-modal attention --- ## Experimental Results | Dataset | F1-Score | AUC | Improvement vs. Baselines ||----------|-----------|-----|-----------------------------|| Operating Room Activity | 88.50 | 91.73 | +3.6% || Surgical Nursing Risk Factors | 88.96 | 92.68 | +4.7% || VRME | 87.73 | 91.04 | +3.5% || OR Workflow Analysis | 88.92 | 91.88 | +3.9% | Refer to Tables 1–4 and Figures 5–8 (pages 13–16) for full benchmark and ablation results. --- ## Future Work - Extend to low-resolution and limited-resource hospital environments. - Introduce self-adaptive learning for cross-institutional generalization. - Incorporate additional modalities such as biosensor or IoT data. --- ## Citation If you use this framework, please cite: @article{wang2025prren, title={Visual Recognition Framework for Operating Room Nursing Risk Factors}, author={Wang, Shanshan}, journal={Frontiers in Computer Science}, year={2025} } --- ## License Released under the **MIT License**. You may use, modify, and distribute this work for research and educational purposes. --- ## Contact **Author:** Shanshan Wang **Email:** [e-mail]