Real-time AI architectures for intraoperative surgical workflow: from temporal modeling and multi-modal fusion to efficient edge deployment

The integration of artificial intelligence (AI) into intraoperative surgical environments represents a transformative paradigm shift in modern surgery. Real-time surgical workflow recognition systems leverage sophisticated temporal modeling architectures, multi-modal data fusion strategies, and efficient edge deployment techniques to provide context-aware decision support during surgical procedures. This narrative review synthesizes contemporary advances in deep learning architectures for surgical video analysis, examining the evolution from recurrent neural networks to transformer-based models, the integration of heterogeneous data sources including video, kinematics, and sensor data, and the computational optimization strategies enabling deployment on resource-constrained edge devices. We critically analyze benchmark datasets, evaluate state-of-the-art performance metrics, and discuss the regulatory landscape governing AI-enabled surgical devices. The review identifies key challenges including context-dependent latency requirements—markedly stricter for robotic teleoperation control loops than for on-stream model inference, which need only keep pace with the ~25 fps surgical-video acquisition rate—together with robust generalization across surgical variations and standardized evaluation protocols. Emerging trends including foundation models for surgical video analysis and vision-language models for automated documentation are discussed, alongside future directions for clinical translation. Clinical trial registration Not applicable.