Cochlear implants represent a critical intervention for individuals with severe-to-profound sensorineural hearing loss. However, surgical insertion of the electrode array remains a challenging procedure where suboptimal trajectories can result in intracochlear trauma, scala translocation, and compromised auditory outcomes.

This paper presents a novel methodology that integrates real-time impedance feedback with spatiotemporal deep learning models to optimize electrode insertion trajectories during cochlear implant surgery. We developed a Spatiotemporal Graph Neural Network (ST-GNN) trained on COMSOL Multiphysics simulation data representing impedance responses across various electrode positions and cochlear anatomies. To enhance robustness and clinical safety, the proposed hybrid system combines the ST-GNN with a classical Decision Tree Classifier.

Our experimental results demonstrate a trajectory prediction accuracy of 94.6%, with trauma risk classification achieving an F1-score of 0.909. The system maintains spatial precision within 0.28 mm depth deviation and 1.42° angular deviation from optimal paths, meeting clinical tolerance requirements. The proposed ST-GNN achieved an MSE of 0.041, representing a 29.3% improvement over the next-best-performing method (GraphSAGE). The hybrid system further reduced the MSE to 0.037, demonstrating the effectiveness of classical AI integration.

Clinical validation scenarios show a 96.3% trauma prevention rate. Real-time inference capabilities with an average latency of 12.3 ms support intraoperative deployment. This approach provides a scalable, autonomous tool for real-time surgical guidance that could significantly improve cochlear implant outcomes.