A Fresh Perspective on Examining Population Emotional Well-Being Trends by Internet Search Engine: An Emerging Composite Anxiety and Depression Index

China has experienced rapid economic development and tremendous social transformation in recent years. Such rapid social change can affect individual emotions, such as increasing insecurity, which is reflected in a marked increase in the incidence of mental illness [1,2]. A nationally representative study of Chinese residents found that 20.4% of people aged 18 and over suffered from anxiety, depression, or both [3]. Concurrently, anxiety is a common mental disorder ranked by the World Health Organization (WHO) as the ninth leading cause of health-related disability, accounting for 3.3% of the global disease burden [4]. Meanwhile, anxiety often co-exists with other mental disorders, especially depression [4]. The number of people with anxiety and depression disorders worldwide was about 246 million and 374 million by 2020, respectively [5].

The group-wide changes in anxiety and depression reflect the psychological characteristics of different regions, which provide a new dimension for the study of social situations [6]. Additionally, different cities have regional specificity, reflected in their development level, medical resources, Internet popularity, and other fields [2,7,8]. These urban characteristics, such as economic development and medical burden, are associated with negative public mental health [2]. Therefore, it is necessary to characterize group sentiment by city to make the analysis more nationally representative of the Chinese region [9].

Traditional assessments of anxiety and depression are facing challenges and difficulties when informing the trends of group psychological characteristics. These assessments are usually made using questionnaires or hospital tests, which require a large workforce and resources [10,11]. Also, there is widespread stigma associated with mental illnesses, such as anxiety and depression [12], significantly increasing the difficulty of detection and limiting its scope. The availability of large-scale mental health surveys in China is limited. The most representative survey, the China Mental Health Survey (CMHS), was conducted only in 2012–2015 [13,14]. Fortunately, with the popularization of the Internet and communication technology, more and more people are using location-based online media services (e.g., Twitter, Weibo) and search engines in their daily lives and posting emotionally-rich texts openly [15,16]. A growing body of research is crunching information from social media to extract specific and reliable emotions from people [17]. It has been confirmed that utilizing social media data from Twitter is effective in monitoring mental health [18]. In addition, Internet search trends are extensively employed in communication, medicine, health, business, and economic research. For example, using Google Search Trends to understanding mental health conditions or explore the correlations between COVID-19 vaccination choices and public mental health [19,20]. In another study, researchers [21] collected Google trends for five words, including “Depression”, “Insomnia”, “Autism”, “Psychologist” and “Psychiatrist” from 25 September 2016 to 19 September 2021 and investigated the future trajectory of public mental health. However, at present, there is no composite sentiment indicator that can reflect the anxiety and depression in China.

Analyzing unstructured emotionally enriched texts from location-based online media services or search engines is challenging. Fortunately, the rapid development of machine learning and text mining technologies provides a new platform for large-scale text mining and online data integration, helping to identify and analyze human behavior and group emotions [16,22]. Recent research proposed a model based on Skip-gram and CNN-BiLSTM for rapid analysis of Sina microblog emotion [23]. In addition, with the key dimensions of the Big Five model, including Openness to Experience, Conscientiousness, Extraversion, Agreeableness and Neuroticism, which are widely accepted as an adequate basis for the representation of human personality. In total, six different computational models were used to compare the ability to identify personality traits based on Facebook texts [24,25]. A system review has pointed out that natural language processing and machine learning have been widely applied in suicide research [26].

The CADI can reflect public mental health at temporal and geographical dimensions. At the temporal level, the CADI indicated an overall increasing trend in negative mental health levels across various regions of China from 2016 to 2021. At the geographical level, the CADI demonstrated regional disparities in public mental health, with more prominent negative mental health observed in the northwest border areas and the southeast coastal regions. A representative cross-sectional epidemiological study of mental health in China revealed a continuous rise in public psychological stress levels over the past several decades [13]. It was noticed that a more severe psychological state in border regions [38]. Additionally, higher rates of mental disorders were observed among both adults and children from prosperous coastal areas [39]. The characteristics of public mental health obtained from the CADI are consistent with previous empirical studies. In addition, there is a significant positive correlation between CADI and the prevalence of mental illnesses. Therefore, it is valid to employ the CADI as a measure of public negative mental health.

We found higher levels of negative mental health in both relatively developed areas (e.g., the eastern coastal areas) and less developed areas (e.g., southwest areas). At the national level, the CADI showed a U-shaped distribution with the GDP, and our findings were consistent with the previous literature. An analysis of more than 450,000 samples found that low income was associated with lower life evaluations and lower emotional well-being, which increased as income increased. However, after reaching a certain income level in the current year ($75,000 in the US), life happiness did not increase further, and the relationship between the two showed a strong convex function [40].

Using GDP stratification can reduce the influence of socioeconomic status and its accompanying factors (e.g., health disparities and living standards) on the association results [36,37]. ACP showed a significant negative correlation with the CADI in both the full and hierarchical models of linear and quantile regression (p 41,42]. On the other hand, with the increased availability of CMIs, people receive more medical financial protection. Still, the actual increase in expenditure and economic burden may be due to patients being treated more frequently in higher-grade hospitals [43].

In addition, compared with the average level, the change of negative sentiment of groups in low GDP areas is more sensitive to the shift in Disposable Income and the Average Cost of Patients. This suggests that poorer areas are more concerned about income levels and healthcare burdens. The lowest income groups have been disproportionately affected by economic expansion. Economic expansion has brought more resources to high-income people than those at the bottom of the income scale, allowing them to adapt to macroeconomic changes [44]. Catastrophic health spending is also concentrated among poorer households [43]. On the other hand, in regions with higher GDP, the unemployment rate shows a significant positive correlation with CADI, which is consistent with previous studies [45]. Unemployment is a stressful life event, involving financial stress and potential benefit deprivation, which is more pronounced in urban areas with higher standards of living and can be accompanied by higher rates of suicide [45,46].

It is noteworthy that the CADI differs significantly from previous studies that solely relied on search engine trends to represent the public mental health of a particular region. Although previous studies have begun to use Internet platforms to construct representative search indicators through the thesaurus and Baidu trend and to evaluate the anxiety situation of each city further [2]. Our Public Opinion Dictionary (POD) was obtained based on the major media of Chinese public opinion. Compared with the mode of specifying the word list in the past, this method guarantees the habits of the times and the characteristics of the population.

Further, we used word correlation and frequency to consider the weight of search terms against target words (“Anxiety” and “Depression”), which reduced the interference of marginal and low-frequency words. On the other hand, with the rapid development of China, the Internet penetration rate of the population in different regions was constantly changing vastly, which was also covered by our CADI model [36]. It is necessary because there may be a correlation between the population size (or the popularity of the Internet) and changes in regional economic levels, search frequency, and even subjective happiness [36,37,38]. Therefore, our CADI has evolved from past methods to be more adaptable and representative.

Although a large sample population and a mature indicator construction process improve the application value of correlation results, our approach still has some limitations. Firstly, our information collection platform covers a relatively limited range of people, which could affect the representativeness and critical stability of the samples concerning age range and whether they are Internet users [47]. Secondly, using the Baidu trends may be affected by “Baidu search suggestions”, a common problem in search-based composite indexes. It’s worth noting that searches related to anxiety and depression may not always be directly associated with individuals experiencing anxiety or depression. Future research should explore the application of CADI to the longitudinal causal framework or predictive analysis. The calculation of CADI across different regions and over a more extended period also requires further validation. However, with the continuous improvement of the network information platform and computer computing power, our Composite Anxiety and Depression Index, construction process, and correlation results will provide extensive reference value.