Page 169 - GHES-3-2

Global Health Economics and
Sustainability
Online health community reviews

maintaining community interaction and emotional health higher the accuracy. However, labeling samples requires a
(Lu et al., 2017). significant amount of manpower. Deep learning methods
Meanwhile, recent studies have further indicated that require large-scale training data. In comparison, the
social media plays an increasingly important role in sharing sentiment dictionary-based analysis method is easier
health experiences and expressing emotions. By analyzing to operate and can improve accuracy by expanding the
user-generated content on platforms like Weibo from sentiment vocabulary according to the text. With higher
emotional word coverage and accuracy, this method yields
spatial and temporal dimensions using machine learning more precise emotional classification. Therefore, this study
and topic modeling, key factors affecting user satisfaction adopted the sentiment dictionary-based method.
can be identified. In addition, medical institutions can
improve their services and response efficiency (Xiang 4.2. Sentiments and concerns
et al., 2023). Several studies utilized composite models
(e.g., combining logistic regression, clustering, and random The vast array of reviews and the multitude of information
available in OHCs significantly influence individuals’
forests) to improve the accuracy of sentiment polarity search for health information, exchange of disease-related
judgment, finding that negative reviews often trigger treatment experiences, and emotional support. Gaining
more interaction and information diffusion. Subdividing insights into the sentiments and concerns expressed in
the functional and emotional elements in doctor-patient the reviews can help users understand the trends within
relationships is of great significance for understanding OHCs, making it a valuable area for exploration. Sentiment
patients’ real needs and optimizing the construction of analysis methods based on sentiment dictionaries provide
OHCs (Chen et al., 2022; Pan et al., 2024; Sun et al., 2022; a reliable and powerful tool for analyzing the sentiment of
Zhang et al., 2014). OHC reviews. The quantification of complex, subjective
Deep learning-based methods have gained prominence text greatly facilitates the study of sentiment orientation
in recent years for sentiment analysis. For example, the and the primary concerns reflected in the reviews.
BERT pre-training language model has replaced Word2Vec According to the program feedback results, positive
and GloVe in embedding word vectors into other models to reviews constituted the highest proportion at 80.3%,
enhance sentiment classification performance (Fang et al., followed by neutral reviews at 11.1%, whereas negative
2020). Another approach involves using a convolutional reviews made up the lowest proportion at 8.6%. In
neural network (MF-CNN) model that leverages diverse general, the number of reviews decreases as the sentiment
feature information. By utilizing abstract features of value calculated by the program decreases, revealing an
word diversity and two methods for computing the extremely skewed distribution within the sentiment value
network input matrix, this model optimizes the sentiment range (0 – 1). Among the three typical diseases – diabetes,
classification effect (Cai et al., 2019). In addition, a leukemia, and depression – diabetic patients gave relatively
computational framework employing a deep learning- fewer positive reviews and more negative reviews, whereas
based language model has been developed for sentiment leukemia patients gave more positive reviews and fewer
analysis using a delayed recurrent neural network (d-RNN) negative reviews. This observation suggests that patients
and its hierarchical variant (Hd-RNN). Experimental have varying emotional orientations toward different
results demonstrate that Hd-RNN outperforms other diseases, and these differences may be related to the
technologies (Chaudhuri, 2022). Furthermore, an LSTM prognosis and treatment of the diseases.
model, named PosATT-LSTM, has been introduced,
which considers the importance of each contextual term Regarding the different sentiments expressed in the
and incorporates position-aware vectors for aspect-level reviews, patients primarily focus on the doctor’s “attitude”
sentiment classification (Zeng et al., 2019). and “patience,” which can be presumed to be important
criteria for their evaluation of doctors. In the positive
Overall, existing research indicates that the study reviews, patients place significant attention on doctors’
of OHCs is crucial for gaining insights into user needs medical skills, with most reviews expressing affirmation
and improving platform service quality. Unlike existing and gratitude for the doctors’ expertise. In the negative
research that focuses on overall sentiment trends, reviews, patients highlighted several key issues, including
emotional distribution, or topic mining, the present study the doctor’s response, sense of responsibility, hospital
directly extracts the key elements from patient evaluations/ environment, and the diagnosis and treatment process.
reviews, making the research objectives and audience more A poor experience in any of these aspects may significantly
focused on real patient experiences and needs. Traditional affect patients’ emotions, thereby leading to negative
machine learning methods require a large number of reviews. The main factor influencing neutral reviews is that
labeled corpora; the more training data available, the patients have mixed experiences regarding the doctor’s

Volume 3 Issue 2 (2025) 161 https://doi.org/10.36922/ghes.7052