Box 2 Most important improvements that were made during the iterative development process

– To first perform a sentiment analysis and then create a separate topic model per sentiment and per question, instead of creating one topic model for both sentiments. This led to more specific topics, from which points of improvements could be derived more easily, increasing the interpretability and actionability

– To not only include the negative feedback topics but also the positive ones, in order to obtain more balanced information. This was found to be essential in selecting and prioritizing points of improvement. In addition, the positive topics were seen as motivators for the healthcare team

– To go from a fixed number of topics to an adaptive approach that automatically chooses the optimal number of topics per subject. This increased the completeness

– To add a quantitative dimension to the qualitative output of the topic model, in order to help prioritize aspects of care that need the most attention

– To include n-grams up to three instead of just using 1 g. This increased the interpretability and actionability of the topics