Unveiling Patient Insights and Experiences: An Integrated Approach to Multi-Criteria Topic Modeling and Classification

This paper unveils a pioneering two-phase integrated framework for analyzing patient experience surveys. The first phase employs unsupervised sentiment and multi-criteria topic modeling using BERT and LDA algorithms, while the second phase involves supervised multi-label classification using BERT and RE algorithms. Leveraging 1772 patient comments from Finger Lakes Community Health (FLCH), our approach demonstrates robust performance, facilitated by the model's multi-label capability that allows discernment of multiple topics within each patient comment. The sentiment analysis revealed 59.1% positive, 26.3% negative, 8.4% mixed, and 6.2% neutral comments. The topic analysis identified seven topics including scheduling, care and service quality, experience with staff, call center, prescriptions and labs, waiting time, and other issues. Notably, the care and service quality topic received overwhelmingly positive sentiments (74.6%), while the call center and waiting time topics exhibited high negativity (66.1% and 70.3%, respectively). The text classification models, particularly the BERT-based model, demonstrated strong performance with balanced accuracy ranging from 82% to 97%. The call center topic achieved the highest balanced accuracy (93%) in the RF-based model. The study contributes to healthcare analytics, providing actionable insights for administrators to enhance patient satisfaction and underscores the efficacy of sentiment and topic modeling approaches in understanding diverse patient experiences.