Over the past few decades, the prevalence of chronic diseases has increased significantly, becoming a global public health concern. The World Health Organization has listed allergic diseases as one of the disease types that require priority research and prevention in the 21st century [1]. As a common chronic disease, allergic rhinitis (AR) is a multifactorial disease that is induced by environmental conditions or certain genes [2]. AR not only has a significant impact on individuals’ sleep, social life, and work attendance but also triggers comorbidities such as conjunctivitis, atopic dermatitis, and asthma [3]. Large-scale flow survey data showed that AR currently affects several people in China alone [4] and with an estimated prevalence between 15% and 20% worldwide [5]. The direct and indirect costs associated with the management of AR are also a significant burden on society. For instance, the total cost of AR in Sweden, with a population of 9.5 million, was estimated at €1.3 (US $1.41) billion annually [6]. These unexpectedly high costs could be related to the high prevalence of disease, in combination with the previously often underestimated indirect costs that arise from reduced work efficiency and absenteeism and the potential costs associated with treating AR comorbidities [6].

Currently, there is no cure for AR, and individuals need to avoid the disease risk factors such as exposure to allergens and inhalation irritants [7] during the long self-management process. Therefore, identifying AR risk factors can provide a reference for patients to help reduce the condition in their daily lives [8].

A plethora of studies have been proposed to identify AR risk factors. These studies recruited participants with symptoms of AR and control participants without AR symptoms from a specific age group or a particular geographical area. These studies collected demographic information, lifestyle habits, family history, comorbidities, and residential areas through questionnaires. Subsequently, they used correlation methods to explore the relationship between these data and AR, aiming to identify the risk factors for AR within the specified age group or geographical area [9]. However, these studies have 2 limitations. First, these studies specifically target certain age groups or geographical areas, and questionnaires can only gather data on specific pieces of information. Owing to the constraints of questionnaire surveys, it is challenging to identify potential risk factors that may be present in individuals’ daily lives. As a result, the risk factors identified through survey-based studies have a limited scope and are incomplete. As such, they provide limited insights for a broader patient population. Second, the survey-based approach demands a commitment to long-term investigation and a substantial effort to collect representative responses [10]. In contrast, collecting information from social media platforms can cover large geographical areas at a comparatively low cost [10]. Social media platforms allow users to share experiences and opinions on various topics [11,12], including personal health issues [13]. Over time, highly unstructured and implicit knowledge has been generated in communities where users frequently participate [14,15], which can provide daily health records that are difficult to obtain from traditional questionnaire surveys. Therefore, social media can become a potential source of information for identifying risk factors for diseases such as AR [16].

Text-mining techniques are an effective tool for using voluminous social media data [17]. Some studies have combined social media data analysis to obtain knowledge about disease risk factors [18,19]. However, the abovementioned studies on disease risk factors used only shallow text features such as the number of social media text items and word cooccurrences, which are not conducive to identifying disease risk factors in the context of colloquial and diverse user expressions [20]. In this study, we designed a text-processing framework to automatically identify risk factors from social media data [21]. We used social media comments to construct a natural language processing–based AR risk factor identification method, aiming to tackle the problems of omission and low accuracy in traditional disease-related information identification methods that rely solely on shallow text features such as word frequency.

To be more specific, we developed an AR risk factor identification method that integrates pretrained word embeddings with text convolutional neural networks (CNNs). The Word2vec algorithm has proven to be superior in text vector representation [20]. This is a prediction-based approach that predicts the neighboring words that are most likely to appear within a window size around a center word in a corpus, resulting in high-dimensional vector representations that capture semantic aggregation. As social media users may mention related topics, such as symptoms and treatments, when describing risk factors in their comments, we used a local context window to achieve better semantic aggregation of AR risk factors, a method that has been demonstrated to be effective for such aggregation. In addition, using the Skip-gram model to train word pairs enables the incorporation of word thematic information, thus improving attention to risk factor phrases. The convolutional network can convolve the text in the word vector dimension and extract critical information through the max-pooling layer operation. In addition, this study used a clustering method with review mechanisms to concentrate on a large amount of text that contains risk factors within the observable range, thereby ensuring the usefulness of the content obtained through text mining.

Our main contributions were as follows:

AR has become a major global issue with a substantial increase in its prevalence in recent years. In Europe, the prevalence of AR among Danish adults progressively increased from 19% to 32% over the past 3 decades [22]. Understanding the risk factors, such as genetic, environmental, and lifestyle factors, helps in the management of AR, thus motivating many studies to focus on identifying potential risk factors. These studies are summarized in Table 1. From Table 1, we observed that the previous studies were based on survey methods, including cross-sectional surveys, cohort studies, and case-control studies.

These studies typically recruited participants with symptoms of AR and control participants without AR symptoms from a specific age group or a particular geographical area, collected demographic information through questionnaires, and then conducted correlation analysis, such as logistic regression, to explore the relationship between those metadata and AR [32]. For instance, Gao et al [9] conducted a cross-sectional survey to investigate the prevalence and risk factors of adult self-reported AR in the plain lands and hilly areas of Shenmu City in China and analyzed the differences between regions. The content of the web-based questionnaire included demographic factors, smoking status, the comorbidities of other allergic disorders, family history of allergies, and place of residence. The unconditional logistic regression analysis was used to screen for factors influencing AR. Finally, they found that the prevalence of AR existed in regional differences. Genetic and environmental factors were the important risk factors associated with AR. However, these studies have 2 limitations. First, these studies specifically targeted certain age groups or geographical areas, and questionnaires can only gather data on specific pieces of information. Owing to the constraints of questionnaire surveys, it is challenging to identify potential risk factors that may be present in individuals’ daily lives. As a result, the risk factors identified through survey-based studies have limited scope and are incomplete and they may provide limited insights for a broader patient population. Second, the survey-based approach demands a commitment to long-term investigation and a massive effort to collect representative responses [10].

Social media sites provide a convenient way for users to continuously update their day-to-day activities, which allows large groups of people to create and share information, opinions, and experiences about health conditions through web-based discussion [11]. Hence, social media can be considered a new data source to assess population health. As shown in Table 2, some studies have combined text-mining techniques to classify and summarize voluminous social media data to obtain knowledge about chronic disease risk factors. Zhang and Ram [33] extracted behavioral features from Twitter posts of asthma users using keywords from an existing knowledge base. Griffis et al [34] collected 25,000 tweets containing and not containing diabetes, identified 5000 common words, used logistic regression to determine which common words were high-frequency expressions of diabetes, and finally grouped these high-frequency words using latent Dirichlet allocation to obtain the risk factors for diabetes. Schäfer et al [35] used syntactic analysis to identify portions of risk factors occurring before or after causal terms, grouped these portions using latent Dirichlet allocation, and obtained the risk factors for gastric discomfort. Pradeepa et al [19] performed clustering on stroke-related tweets using the Probability Neural Network, used the Apriori algorithm to identify frequent word sets related to risk, and thus identified risk factors for stroke [19]. In addition to the aforementioned approaches that use shallow text features such as keywords, frequent word sets, high-frequency words, and syntactic features for disease risk factor identification, other studies [36-38] trained risk factor classifiers using machine learning methods such as Naive Bayes, Maximum Entropy Model, and Naive Bayes Classifier–Term Frequency Inverse Document Frequency. These classifiers predict the presence of risk factors in text based on discrete vector representations such as bag-of-words and n-gram.

The current methods for identifying disease risk factors on social media fall into 2 categories: shallow text feature methods and discrete word vector representations. Shallow text feature techniques often fail to capture important risk factors resulting in low accuracy, whereas discrete word vector approaches struggle to keep up with the dynamic vocabulary of social media text, missing new words, and trending expressions, thus inadequately representing the information conveyed.

Natural language processing technology promotes text analysis based on social media comments [39]; this technology can learn the deeper semantic features of the comment text and the features that are consistent with the current context, according to different training corpus, to input a better text vector representation for downstream classification tasks. Some researchers have used large-scale pretrained language models [40], global matrix decomposition [41], and local context windows [42] for text vector representation. Local context windows are more suitable for semantically aggregating AR risk factors [43]. Skip-gram and Continuous Bag-of-Words Model (CBOW) are prediction-based methods that learn the semantic representation of a center word by predicting the most likely neighboring words within a window size in a corpus. When users narrate risk factors in their comments, they may also mention symptoms, treatments, and other topics. These global contexts may dilute the key features of the risk factors expression. CBOW averages the context words to predict the target word and tends to predict high-frequency words in the corpus. In contrast, Skip-gram gives each word a chance to be a center word, making it better at predicting rare words compared with CBOW [44]. Therefore, in situations where social media users express a wide variety of ideas, the Skip-gram model can yield satisfactory outcomes. Moreover, the Skip-gram approach uses word pair training, which facilitates the incorporation of topic information into words [45], resulting in the generation of high-dimensional word vectors that feature semantic aggregation and topic enhancement. Therefore, we selected Skip-gram as the word-embedding model for our study.

Text classification has evolved to deep learning models, mainly including CNN-based models [46], recurrent neural network (RNN)–based models [47], and transformer models [48]. For the CNN algorithm, convolutional networks can convolve text on the word vector dimensions and extract key information through pooling layer operations. Consequently, this algorithm is capable of using essential data for classification tasks. Therefore, we used TextCNN for classifier training and evaluated the performance of RNN and transformer models on this task.

The framework used in this study consisted of 3 parts as shown in Figure 1. The first part was data collection and processing, aimed at obtaining a clean data set. The second part was risk factor identification, which included the proposed TopicS method and training of a risk factor classifier. The implementation steps were as follows: (1) semiautomatically constructing a risk factor topic dictionary, (2) generating high-dimensional word vectors enhanced by TopicS-generated topics, and (3) vectorizing annotated text and training a risk factor classifier. The third part is text clustering and keyword extraction, which uses the ClusterREV method to cluster the identified risk factors and extract keywords from every category.

Zhihu is a Chinese social media platform where people discuss topics in an web-based forum format. In May 2022, the Zhihu subcommunity allergic rhinitis had 1.04 million discussions. The posts on this social media platform allow other users to comment [49], and people can explain their situations to provide support or seek help effectively. Therefore, these comments provide a rich source of data for investigating the risk factors reported by different users [50]. In this study, we trained domain-specific word representations based on experimental data. A relatively domain-specific input corpus [51] is better at extracting meaningful semantic relations than a generic pretrained language model [52]. We crawled all the data from May 2012 to May 2022 under the topic allergic rhinitis on Zhihu, obtaining a total of 9628 posts and 33,747 comments, including the post ID, comment ID, and post and comment content.

In this study, we preprocessed the data through regularization, stop word removal, and word separation. First, we removed special symbols, such as URLs and emoticons, in the comments through regularization and stop word removal to reduce the interference of noise with the text analysis task. Then, we compiled a dictionary of 169 specialized terms, including types of AR, medications, and comorbidities, to reduce the probability of incorrect word segmentation. After word separation, we obtained a lexicon of 68,863 words and ranked the words according to the number of occurrences. We found that the top 10,000 words accounted for 94.83% of the total words, suggesting that many words recurred and a relatively simple word vector could effectively train the model [53]. This further confirms the efficacy of our decision to use Skip-gram as the foundational model.

We observed ultrashort comment noise in the comments (eg, “Thank you!”). It is important to note that these ultrashort comments do not include any personal medical information. The ultrashort comments were filtered, resulting in 33,039 valid comments. This operation can effectively minimize the impact of noise on downstream text classification tasks. Table S1 in Multimedia Appendix 1 presents the examples of valid comments.

The data must be labeled before supervised learning and then trained end to end. If a comment directly mentions an allergen or indicates a condition that leads to the appearance or worsening of symptoms, the comment will be labeled as 1, indicating the presence of risk factors, as shown in Figure 2.

We randomly chose 2030 comments from the 33,039 comments, and 3 researchers labeled each comment as containing or not containing risk factors. To ensure high interannotator consistency, all 3 researchers annotated all 2030 comments. In cases with uncertainty in labeling, the 3 researchers discussed and arrived at a final label. After annotating and eliminating comments with religiously controversial content, 2000 labeled comments remained, consisting of 996 comments containing risk factors and 1004 comments not containing risk factors. The data set was divided into a 90% training set and a 10% test set. The 90% training set was further divided into 10 subsets, with 9 subsets used for training and the remaining subset used for validation, performing 10-fold cross-validation.

We used a combination of manual labeling and similarity calculation to identify keywords related to risk factors. Subsequently, we constructed a table of topic words using a semiautomated approach. The process of constructing the dictionary is depicted in Textbox 1 and is as follows: (1) label 400 randomly selected comments as described in the Annotation section, thereby obtaining 198 comments with risk factors; (2) extract risk factor phrases from annotated comments; (3) obtain risk factors topic word list; (4) remove duplicate word list, and the words in the current topic are used as seed words, word_set; (5) use Skip-gram to find the top similar words to expand the topic words; (6) repeat steps 3 through 5 to expand the topic word; and (7) finally, obtain the topic words for the risk factor. A large weight was assigned to the risk factor theme words. Table S2 in Multimedia Appendix 1 shows examples of the risk factor topic dictionary.

As the use of text data from social media involves user privacy, this study adopted the following steps for deidentification: (1) We removed user account information and retained only anonymous comment information. (2) We used regular expressions to match and delete URLs and email addresses in the comments. (3) During the annotation process, annotators received only text that did not involve personal information. To evaluate the quality of deidentification, we randomly selected 500 text items for manual inspection and did not find any instances containing personal identity information. Our data are sourced from public discussions on Zhihu, a social media platform that can be accessed without registration. We followed strict ethical research protocols similar to the guidelines by Eysenbach and Till [54]. In addition, to protect the anonymity of participants, we have implemented measures including the removal of user information and avoiding verbatim quotations to prevent identification through search engines, protecting the privacy and security of personal data. It should be mentioned that our study was focused on the post level; we do not anticipate any negative ethical impact from our analysis.

TopicS performed 2 tasks during training, as shown in Figure 3. The first task was to predict the neighboring words within the window of the central word. The second task was to predict the topic of the central word; the topic dictionary used for this purpose is described in the Topic Dictionary Construction section.

The specific formula calculations for the loss function design, parameter updates, and error backpropagation of TopicS are explained subsequently.

First, we defined the loss function. For each word in the corpus, we used it as the central word for a sliding operation with a window size of c; let S be the training sequence (w₁,w₂,...,w_T), whereas w_i denotes the ith word in the sequence. The subscript T represents the total number of unique words in the corpus. In addition to predicting the contextual word of the central word, we must also predict the topic score of the central word. Therefore, the loss function comprised 2 parts: L_cont and L_topic, and the overall loss was denoted by L_s. Our training objective was to minimize the loss function:

The initial word vector was represented as one-hot vector. The central word was denoted as w_center, the surrounding words of the central word were denoted as w_(O,cont(i)), and the central word topic information was denoted as w_(O,topic). The weight parameter λ was used to balance the loss of L_cont and L_topic. After sliding the window to browse the entire corpus to average all loss window losses, error backpropagation was performed to update the parameters. The input matrix was represented as the central word vector. The actual loss function can be expressed as

Second, we introduced the rules for updating parameters. The parameters were updated through error backpropagation Here, y_true represents the true value of the contextual words in the window, and y_pred represents the predicted value.

Finally, we can update the word representation.

In this study, we chose TextCNN as the classification model. In the risk factor identification task, some key semantic information is more important, and TextCNN can efficiently use the key information for classification with minimal cost consumption. We represented the manually annotated text as a vector matrix using high-dimensional word vector representations trained by the TopicS model, which aggregates local contextual and topic information and uses it as input for the TextCNN model. Then, the TextCNN algorithm leverages convolutional kernels of different sizes to extract multiple n-gram text features and uses convolutional operations in a fixed window to combine word representations to capture local information. Our input word vector combined the topic information of words, and the most important features in the convolution operation can be extracted using the maximum pooling operation as shown in Figure 4.

The clustering task is to group similar risk factors. In this study, a large amount of text containing risk factors was clustered into a manually observable number of categories, making it easier to comprehend their content. This study enhances the single-pass algorithm and integrates it with a manual review to cluster the risk factors identified in the text classification, ensuring the validity of the clustering results. The main concept of single-pass clustering [55] is to match informational text items based on their similarity values without the need to determine the number of clusters in advance. This makes it suitable for clustering tasks with an unknown number of clusters. However, traditional single-pass clustering uses only one-loop traversal, which may result in previously entered text items completing the traversal earlier. This can cause their similarity to the previous topics to be slightly lower than the threshold and lead to them being recreated as new categories, ultimately affecting the clustering effect.

As shown in Figure 5, we improved the single-pass algorithm by retraversing the categories that were clustered separately after all the text items had been traversed to handle any missed text. After the automated clustering was completed, we conducted a manual review to ensure the reliability of the clustering.

Moreover, this study uses a keyword cloud visualization of category content to quickly understand the themes and characteristics of each cluster and compare the differences between different clusters. TextRank [56] was selected to extract category keywords, which considers only the voting scores of words in a single document; common words that frequently appear in a single document easily obtain high scores [57]. We treated each category as a single document for keyword extraction. As risk factors appear more frequently in categories, TextRank can effectively extract risk factors and surrounding words, preserving category content information as much as possible and reflecting the true content of the risk factors.

In this section, we present the performance of the classifier and the findings based on the categorization of all the comments in the clean data set using the classifier. Our approach involved visualizing the clustering results of the risk factors to comprehend the primary elements of these factors. We also explored the pathogenic mechanisms associated with these risk factors.

We used standard text-mining evaluation metrics such as accuracy, precision, recall, and F₁-score to evaluate the performance. Precision assesses how many risk factors the model identifies correctly, and recall measures how many risk factors the model can identify on its test set. As we aimed to identify as many AR risk factors as possible to provide comprehensive references for individuals, recall was more important than precision in our study.

We set 7-word embedding dimensions ranging from 100 to 400. Table 3 displays the classification results of the TextCNN classification model with the 7 dimensions of Skip-gram and TopicS word vectors. In addition, TextRNN and transformer models were evaluated with the 7-word embedding dimensions of TopicS or Skip-gram, as shown in Tables S3 and S4 in Multimedia Appendix 1; the classification models performed better when the word-embedding dimension was 100 or 150, as shown in Table 4, which includes the results with best-performing dimensions. This study conducted word representation learning on a domain-specific input corpus, where low dimensionality was found to be sufficient to represent the features of the corpus [58]. Moreover, TopicS not only improved precision but also significantly increased recall for all 3 models, as shown in Table 4.

Table 4 shows that TextCNN has the highest accuracy and recall rate among the 3 classification models. The highest accuracy achieved by our classification model was 0.9594, which used a 150-dimension word-embedding representation obtained from TopicS. In other words, TextCNN can detect more risk factors and minimize the loss of risk factors resulting from classification errors. The CNN model can extract key information similar to n-grams in sentences. The combination of TopicS and TextCNN can enhance topic information and achieve an aggregation effect. Our implementation process was the simplest and consumed the least resources. Our model examined 30,372 comments and identified 5221 comments containing risk factors.

We clustered the text items obtained from the text classification into 28 categories and extracted keywords from each category to better understand the content. Table 5 shows the top 5 categories and their corresponding keywords. The complete list can be found in Table S5 in Multimedia Appendix 1. We used category 1 as an example to explain the category formation process and demonstrate the validity of the qualitative results. As shown in Table 4, we labeled category 1 as Season based on the analysis of keyword weights and relative comments. The comments related to this category focused on seasonally induced AR, with factors such as changes in the weather during seasonal transitions and colder temperatures during winter, which can exacerbate symptoms. We also counted the number of text items in each category and found that seasonal, regional, mites, and weather changes were common risk factors for most patients. In addition, patients’ unhealthy lifestyle habits were also important risk factors widely present in research investigations. Furthermore, most patients reported experiencing symptoms at specific times (eg, “morning”), but researchers have paid little attention to the timing of symptom occurrence (which we refer to as time points).

We referred to the relevant literature on the risk factors associated with AR to confirm whether the extracted risk factors were consistent with the general medical consensus. Our findings are novel compared with those in the literature [59]. Previous survey-based studies have explored only the correlation between risk factors and AR, whereas our experimental data provide insight into the potential pathogenesis of reported risk factors. The following section provides a theoretical discussion of potential pathways for several risk factors that trigger AR:

This theoretical discussion regarding the potential pathway of risk factors that trigger AR can guide the development of detailed AR intervention measures. For example, patients with AR can pay attention to pollen concentration and temperature changes and adjust their outings and clothing accordingly based on the characteristics of the season; they can set the air conditioner to turn on or off based on their waking time to reduce the inhalation of cold air when waking up. Furthermore, they can adjust their sleeping position to reduce the frequency of nighttime symptoms.

This study aimed to identify the risk factors for AR based on social media comments. To do so, a data set of comments related to AR was collected, processed, and analyzed. The data set covered a consecutive period from May 2012 to May 2022. Overall, this analysis provided new insights into three main questions: (1) How many comments contained AR risk factor information? (2) How many categories can these risk factors be summarized into? (3) How do these risk factors trigger AR?

In assessing the identification of AR risk factors, we found that TopicS enhanced both precision and recall. TextCNN outperformed other models, achieving an accuracy of 0.9594 with a 150-dimension TopicS embedding. Analyzing 30,372 comments, our model pinpointed 5221 comments with risk factors. Categorizing the text items led to 28 distinct categories, with seasonal factors, regional variations, mites, weather changes, and unhealthy lifestyle habits emerging as common risks.

Furthermore, our research into AR risk factors revealed how risk factors trigger AR and uncovered the frequently reported, but underresearched, risk factors by affected individuals. Seasonal changes, especially during spring and autumn, increase exposure to pollen allergens, with varying climatic conditions affecting the development of allergies. Poor habits, such as smoking, irregular sleep, and frequent use of air conditioning, compromise immunity and heighten AR susceptibility. Dust mites, influenced by humidity, stand out as a primary allergen, with food items and indoor factors, such as animal dander, also triggering AR. Industrial pollutants and outdoor environmental factors amplify AR risk. Notably, AR symptoms intensify during mornings and evenings, which is likely influenced by circadian rhythms and environmental factors.

This study has some limitations. Our study was based on the self-reported nature of social media data, and the lack of more detailed information from the study participants was a concern. Our statistics showed that seasonal factors, regional variations, mites, weather changes, and unhealthy lifestyle habits emerge as common risk factors, which is consistent with the findings of other studies based on surveys. Although social media may lack in-depth patient information, it provides an effective method of collecting breadth of data. Social media data can be gathered 24 hours a day and are an extremely efficient way to rapidly update new knowledge into the risk factor knowledge base. In the future, our framework can be expanded in 2 ways. First, the framework can track the development trends and changes in AR risk factors by leveraging real-time internet data sets. Second, the framework can be generalized and extended to detect patterns, trends, and risk factors for other chronic diseases such as type 2 diabetes.

In this model improvement study, we proposed a topic-enhanced word-embedding model to improve the accuracy and recall of the text classification, namely to uncover less common or other types of risk factors based on social media data that have not been previously reported. The risk factors identified in this study can be a helpful reference for people with AR to reduce the development of the disease in their daily lives. This study establishes a knowledge base of potential risk factors for individuals who may not be aware of the factors that could trigger their symptoms. Patients can compare their lifestyle habits and medical history to identify their risk factors, which could help reduce the frequency of episodes and prevent the decline in their quality of life caused by blindly avoiding potential triggers. Our findings demonstrate the practicality and feasibility of using social media data for investigating disease knowledge. These findings may provide guidance for the development of management plans and interventions for AR.