Abstract
PURPOSE: The automatic extraction of knowledge about intervention execution from surgical manuals would be of the utmost importance to develop expert surgical systems and assistants. In this work we assess the feasibility of automatically identifying the sentences of a surgical intervention text containing procedural information, a subtask of the broader goal of extracting intervention workflows from surgical manuals. METHODS: We frame the problem as a binary classification task. We first introduce a new public dataset of 1958 sentences from robotic surgery texts, manually annotated as procedural or non-procedural. We then apply different classification methods, from classical machine learning algorithms, to more recent neural-network approaches and classification methods exploiting transformers (e.g., BERT, ClinicalBERT). We also analyze the benefits of applying balancing techniques to the dataset. RESULTS: The architectures based on neural-networks fed with FastText's embeddings and the one based on ClinicalBERT outperform all the tested methods, empirically confirming the feasibility of the task. Adopting balancing techniques does not lead to substantial improvements in classification. CONCLUSION: This is the first work experimenting with machine / deep learning algorithms for automatically identifying procedural sentences in surgical texts. It also introduces the first public dataset that can be used for benchmarking different classification methods for the task.