Mining topic and sentiment dynamics in physician rating websites during the early wave of the COVID-19 pandemic: Machine learning approach

Abstract

Introduction: An increasing number of patients are voicing their opinions and expectations about the quality of care in online forums and on physician rating websites (PRWs). This paper analyzes patient online reviews (PORs) to identify emerging and fading topics and sentiment trends in PRWs during the early stage of the COVID-19 outbreak.

Methods: Text data were collected, including 55,612 PORs of 3430 doctors from three popular PRWs in the United States (RateMDs, HealthGrades, and Vitals) from March 01 to June 27, 2020. An improved latent Dirichlet allocation (LDA)-based topic modeling (topic coherence-based LDA [TCLDA]), manual annotation, and sentiment analysis tool were applied to extract a suitable number of topics, generate corresponding keywords, assign topic names, and determine trends in the extracted topics and specific emotions.

Results: According to the coherence value and manual annotation, the identified taxonomy includes 30 topics across high-rank and low-rank disease categories. The emerging topics in PRWs focus mainly on themes such as treatment experience, policy implementation regarding epidemic control measures, individuals' attitudes toward the pandemic, and mental health across high-rank diseases. In contrast, the treatment process and experience during COVID-19, awareness and COVID-19 control measures, and COVID-19 deaths, fear, and stress were the most popular themes for low-rank diseases. Panic buying and daily life impact, treatment processes, and bedside manner were the fading themes across high-rank diseases. In contrast, provider attitude toward patients during the pandemic, detection at public transportation, passenger, travel bans and warnings, and materials supplies and society support during COVID-19 were the most fading themes across low-rank diseases. Regarding sentiment analysis, negative emotions (fear, anger, and sadness) prevail during the early wave of the COVID-19.

Conclusion: Mining topic dynamics and sentiment trends in PRWs may provide valuable knowledge of patients' opinions during the COVID-19 crisis. Policymakers should consider these PORs and develop global healthcare policies and surveillance systems through monitoring PRWs. The findings of this study identify research gaps in the areas of e-health and text mining and offer future research directions.

Keywords: COVID-19; Discrete emotions; Dynamics of healthcare topics; LDA; Text mining; Topic modeling.

Fig. 1

Overall proposed methodology.

Fig. 2

Coherence score against each topic number.

Fig. 3

Conceptual diagram of topic life cycle and two topic evolution processes.

Fig. 4

Number of new cases, deaths, and online reviews during the study period.

Fig. 5

Distance map between topics (high-rank and low-rank diseases) PC: Principal component.

Fig. 6

Top 10 most relevant terms for Topic 1 in the high-rank disease category (12.9 % of tokens).

Fig. 7

The dynamics of 15 topics across high-rank diseases.

Fig. 8

The dynamics of 15 topics across low-rank diseases.

Fig. 9

Overall emotion trends during the early wave of the COVID-19 pandemic.