Abstract
Introduction: Manually analysing the quality of large amounts of written feedback comments is time-consuming and demands extensive resources and human effort. Therefore, this study aimed to explore whether a state-of-the-art large language model (LLM) could be fine-tuned to identify the presence of four literature-derived feedback quality criteria (performance, judgment, elaboration and improvement) and the seven CanMEDS roles (Medical Expert, Communicator, Collaborator, Leader, Health Advocate, Scholar and Professional) in written feedback comments.
Methods: A set of 2,349 labelled feedback comments of five healthcare educational programs in Flanders (Belgium) (specialistic medicine, general practice, midwifery, speech therapy and occupational therapy) was split into 12,452 sentences to create two datasets for the machine learning analysis. The Dutch BERT models BERTje and RobBERT were used to train four multiclass-multilabel classification models: two to identify the four feedback quality criteria and two to identify the seven CanMEDS roles.
Results: The classification models trained with BERTje and RobBERT to predict the presence of the four feedback quality criteria attained macro average F1-scores of 0.73 and 0.76, respectively. The F1-score of the model predicting the presence of the CanMEDS roles trained with BERTje was 0.71 and 0.72 with RobBERT.
Discussion: The results showed that a state-of-the-art LLM is able to identify the presence of the four feedback quality criteria and the CanMEDS roles in written feedback comments. This implies that the quality analysis of written feedback comments can be automated using an LLM, leading to savings of time and resources.