Characterization, Extraction, and Impact Assessment of Semantic Aspects in Online Physician Reviews

Abstract

Online physician reviews have become an important source of information influencing patients' decisions regarding their physician selection. However, the rapid growth in the number of physician reviews makes them difficult to use in support of patient decision-making. Previous work on physician reviews has primarily focused on examining ratings but ignored important aspects embedded in text content. The main goal of this dissertations is to analyze and extract the semantic aspects of care from physician reviews to facilitate the assessment of physician service quality. The dissertations mainly consists of three inter-related topics: characterization of the semantic aspects in physician reviews, assessment of the aspects' impact on patient satisfaction, and automatic extraction of the semantic aspects from the reviews. First, we develop a taxonomy for the semantic aspects of care reported in physician reviews using text mining and content analysis methods. We validate the taxonomy with the Consumer Assessment of Healthcare Providers and Systems (CAHPS) survey. The evaluation shows that the proposed taxonomy not only covers all dimensions in the CAHPS survey but also contains additional aspects that could help improve our understanding of patients' experiences. We further examine the effects of medical specialties and review platforms on the semantic aspects of physician reviews. Second, drawing on the aspects extracted from physician reviews, we investigate their impact on patient satisfaction using logistic regression models. The analysis results uncover significant relations and important factors affecting patient satisfaction with physician services. Third, we develop an effective framework for extracting semantic aspects from physician reviews using state-of-the-art deep learning techniques. We use a semi-supervised autoencoder that leverages an attention mechanism to automatically detect aspects from an unlabeled corpus of physician reviews with a few seed words. To address the problem of rare aspects in physician reviews, we propose a novel sampling method that oversamples the underrepresented aspects with semantically relevant and linguistically diverse samples by applying the k-means clustering method. We compare the performance of the proposed method with other text oversampling techniques for aspect classification leveraging the Bidirectional Encoder Representations from Transformers (BERT). Our experimental results demonstrate that the proposed sampling method leads to significant improvement in classification performance over conventional sampling methods. The findings of this dissertations research can help improve the efficacy of physician reviews for health consumers. They offer an alternative venue for assessing health service quality by extracting, identifying, and examining important aspects of care from physician reviews. They can also help healthcare providers and policymakers better understand and meet patients' needs to improve service quality.