Weakly supervised text classification on free-text comments in patient-reported outcome measures.

Abstract

Background: Free-text comments in patient-reported outcome measures (PROMs) data provide insights into health-related quality of life (HRQoL). However, these comments are typically analysed using manual methods, such as content analysis, which is labour-intensive and time-consuming. Machine learning analysis methods are largely unsupervised, necessitating post-analysis interpretation. Weakly supervised text classification (WSTC) can be a valuable analytical method of analysis for classifying domain-specific text data, especially when limited labelled data are available. In this paper, we applied five WSTC techniques to PROMs comment data to explore the extent to which they can be used to identify HRQoL themes reported by patients with prostate and colorectal cancer.

Methods: The main HRQoL themes and associated keywords were identified from a scoping review. They were used to classify PROMs comments with these themes from two national PROMs datasets: colorectal cancer (n = 5,634) and prostate cancer (n = 59,768). Classification was done using five keyword-based WSTC methods (anchored CorEx, BERTopic, Guided LDA, WeSTClass, and X-Class). To evaluate these methods, we assessed the overall performance of the methods and by theme. Domain experts reviewed the interpretability of the methods using the keywords extracted from the methods during training.

Results: Based on the 12 papers identified in the scoping review, we determined six main themes and corresponding keywords to label PROMs comments using WSTC methods. These themes were: Comorbidities, Daily Life, Health Pathways and Services, Physical Function, Psychological and Emotional Function, and Social Function. The performance of the methods varied across themes and between the datasets. While the best-performing model for both datasets, CorEx, attained weighted F1 scores of 0.57 (colorectal cancer) and 0.61 (prostate cancer), methods achieved an F1 score of up to 0.92 (Social Function) on individual themes. By evaluating the keywords extracted from the trained models, we saw that the methods that can utilise expert-driven seed terms and extrapolate based on limited data performed the best.

Conclusions: Overall, evaluating these WSTC methods provided insight into their applicability for analysing PROMs comments. Evaluating the classification performance illustrated the potential and limitations of keyword-based WSTC in labelling PROMs comments when labelled data are limited.