Fig. 6. Processing steps involved in the discovery of topics from surgical feedback instances.

These steps are provided under the BERTopic framework¹⁹. Step 1 involves obtaining numerical representations of feedback texts using pretrained embeddings. Step 2 applies dimensionality reduction and projection of the mebeddings into lower-dimensional space. Step 3 clusters the feedback instances into topics based on similarity metric for text represenations. Step 4 summarizes the contents of each topic cluster using concise titles provided by GPT-4 prompting.