Operative notes form a vital part of patient management. However, their manual preparation is vulnerable to human mistakes, particularly under the stress of busy clinical settings. Automatically generating these notes from surgical video recordings can reduce administrative demands, increase reliability, and provide richer contextual detail. For endoscopic pituitary surgery, experts reached consensus on 27 distinct steps. In the 97 videos collected for this research, an experienced surgeon annotated the precise timing of each step. A three-stage model was designed to automatically detect the presence of any given step within a video. The first stage applied a convolutional neural network to classify each video frame in binary fashion for every step. The second stage used a discriminator to convert these frame-level results into a binary decision about whether each step was present in the entire video. The third stage utilized an accumulator to combine these video-level decisions into a multi-label classification of all steps present in the recording. Trained on 77 videos and assessed on a held-out set of 20 videos, the system achieved a weighted F1 score of 0.80. The derived classifications were inserted into a standardized clinical template and further augmented with quantitative video analytics. This investigation therefore confirms that automatic production of operative notes from surgical videos is achievable and can meaningfully support surgeons in their documentation responsibilities.
