Response to ‘Methodological considerations on artificial intelligence‐based phase recognition in robot‐assisted radical prostatectomy’

We thank Bao et al. [1] for their thoughtful letter regarding our study on artificial intelligence (AI)-based phase recognition in robot-assisted radical prostatectomy (RARP) [2]. We welcome this constructive dialogue, which highlights important methodological considerations for the advancing field of surgical AI. We address each point below. Inter-annotator agreement: we agree that reporting formal inter-annotator agreement metrics, such as Cohen's kappa coefficient, represents a best practice in surgical video annotation research. In our study, surgical phases were defined by objective, platform-independent anatomical landmarks, and annotations were performed by two board-certified urologists using a collaborative consensus approach. While this design minimised classification ambiguity, we acknowledge that reporting quantitative agreement from initial independent annotations would further enhance transparency. We have incorporated this as a standard element in our ongoing prospective studies. Temporal modelling: we appreciate the emphasis on integrating temporal features into surgical phase recognition. As we discussed in our original manuscript, the selection of EfficientNet with frame-wise classification was an intentional design decision to enable gradient-weighted class activation mapping (Grad-CAM) visualisation, providing clinicians with interpretable insights into model decision-making. A recent systematic review has confirmed that hybrid approaches combining spatial feature extraction with temporal sequence modelling consistently outperform frame-based architectures alone [3]. We concur that this represents a key frontier for improving recognition accuracy, particularly in phases with similar visual appearances but distinct procedural sequences, and we are actively exploring such hybrid approaches in our ongoing work. Clinical applicability in complex scenarios: our dataset comprised all consecutive cases during the study period, naturally encompassing varying degrees of surgical complexity. However, we agree that stratified analysis of model performance across specific intraoperative events—such as significant bleeding or anatomical variations—is essential prior to clinical deployment. This represents a critical validation step that we are addressing in our current multicentre study with expanded case annotation. Automated surgical phase recognition serves as a foundational technology enabling phase-specific data collection and objective skill assessment, as demonstrated by a recent multicentre study [4]. The methodological considerations raised by Bao et al. [1] will be valuable in guiding the standards for these emerging applications. We believe that the points raised by Bao et al. [1] and our shared commitment to methodological rigour will help establish robust standards for surgical AI research, ultimately benefiting patient care through safer and more effective robotic surgery. The authors declare no conflicts of interest related to this work.