Principal Findings
Our study, conducted from January 2017 to December 2021, revealed significant correlations between the mean of the total weekly RSVs, including diarrhea and its synonyms, and total weekly ED visits for diarrhea-related symptoms. The trend in mean total weekly RSVs closely mirrored that of the total weekly ED visits. Notably, the correlations between weekly ED visits and weekly RSVs were stronger for synonyms such as “upset stomach,” “watery diarrhea,” and “acute enteritis” than for the general term “diarrhea.” The weekly RSV for “upset stomach” was significantly correlated with weekly ED visits for diarrhea at lags of 1 and 2 weeks. Similarly, the weekly RSVs for “watery diarrhea” and “acute enteritis” were significantly correlated with ED visits for watery diarrhea at lags of 1 and 2 weeks. However, weekly ED visits with vomiting and diarrhea showed no significant relationship with weekly RSVs, except those with “upset stomach.” ARIMAX models further confirmed the lag times between weekly RSVs for diarrhea synonyms and weekly ED visits, demonstrating strong predictive performance across weekly RSVs and lag times. Along with that, search engine users older than 50 years showed the highest proportion of both RSVs and ED visits for diarrhea-related symptoms. Among the search terms, “diarrhea” and “upset stomach” each accounted for 26% of the total search volume. In comparison, ED visits for acute and watery diarrhea accounted for 46% and 40%, respectively.
Many people prefer to search the internet for health information before visiting EDs [34]. A previous study has shown that the relative frequency of searches for gastrointestinal symptoms, including diarrhea, closely mirrored the changing incidences of cases in large inpatient datasets [17]. Our findings align with those of previous studies, indicating that search trends for diarrhea symptoms reflect ED visit trends. Google Trends automatically compiles search terms related to each symptom and other related search terms; however, the exact compilation process and methodology are not publicly disclosed. Studies in Korea have shown that representative keywords for foodborne diseases, such as “Seol-sa” (diarrhea) [9], can track population-level changes in the incidence of diarrhea through search volumes [18]. This finding supports the idea that search engine data can be used to monitor ED visits.
Our results showed that weekly ED visits for diarrhea had a higher correlation with weekly RSVs for diarrhea synonyms, including “upset stomach”, “watery diarrhea”, and “acute enteritis”, than with those for “diarrhea”. Adults suffering from diarrhea, particularly acute forms, commonly seek medical evaluation in EDs [35]. The criteria for acute diarrhea include volume depletion and 6 or more stools in 24 hours, making “watery diarrhea” a more relevant term for acute diarrhea than chronic diarrhea [36,37]. This explains why weekly RSVs for “watery diarrhea” are related to weekly ED visits for diarrhea-related symptoms, as acute diarrhea requires immediate medical evaluation [38]. Similarly, “upset stomach” and “acute enteritis” can be the terms used for acute diarrhea, supporting the observation that searches for these keywords precede ED visits. In addition, acute diarrhea lasts for less than 14 days, while diarrhea lasting more than 14 days is termed “persistent”, and those lasting over 1 month are termed “chronic” [30]. While vomiting episodes are typically short in duration, they do not correlate well with internet searches conducted 1–2 weeks before ED visits. Therefore, our results showed that weekly RSVs indicating acute diarrhea became notable within 1–2 weeks, whereas weekly ED visits for vomiting with diarrhea did not show a significant relationship with weekly RSVs for diarrhea and its synonyms. In addition, the weak or nonsignificant correlations between weekly ED visits for vomiting with diarrhea, and weekly RSV for diarrhea and its synonyms may reflect a discrepancy between actual symptom presentations in ED visits and real-world search behaviors. While patients who visit the ED with “vomiting with diarrhea” are experiencing both symptoms concurrently, most internet users tend to search for only one dominant symptom at a time rather than enter queries that combine multiple symptoms [39]. As a result, compound search phrases that include both vomiting and diarrhea are rarely used, which may lead to lower relative search volumes and weaker correlations with ED visit data, despite their clinical relevance.
A previous study conducted in Germany reported that individuals aged 36-55 years were the most active users of search engines for health-related information, while those older than 56 years demonstrated the lowest levels of engagement [40]. In contrast, South Korea exhibits one of the highest internet penetration rates globally, including among older adults, with more than 90% of individuals aged 50 years and older using the internet as of 2021. Recent national statistics show that 70.2% of Korean internet users search for health and medical information, with comparable rates among men (70.7%) and women (69.7%). Age-stratified data further indicate that individuals in their 30s had the highest rate of health information-seeking (80.3%), followed closely by those in their 40s (79.9%), 50s (78%), and 60s (75.7%) [41]. This upward trend in digital engagement among older adults is supported by previous research showing a narrowing digital health divide and increasing internet use for health-related purposes in older populations [42,43]. In the context of our study, individuals older than 50 years not only exhibited the highest frequency of diarrhea-related ED visits but also demonstrated the highest levels of related web-based search activity. This may be attributable to 2 converging factors: the higher symptom burden in this age group and their growing reliance on internet-based health resources. Whereas younger adults are more inclined to seek health information through social media platforms [42], older adults tend to prefer structured, one-way sources such as search engines, further reinforcing their visibility in search volume data.
The overlap of our study period with the early phase of the COVID-19 pandemic (2020-2021) may have influenced both internet search behavior and ED usage patterns. During this time, public awareness of gastrointestinal symptoms, some of which overlap with COVID-19 manifestations, may be heightened due to widespread media coverage and health messaging. This may have led to increased search volume for symptom-related terms such as “diarrhea” or “upset stomach,” even in the absence of actual illness, thereby introducing potential noise into search trends [44]. Furthermore, studies have shown that many individuals delayed or avoided ED visits for non–COVID-19 conditions due to fear of infection or changes in health care access during the pandemic [45,46]. These behavioral shifts may have altered the typical relationship between symptom occurrence, web-based search activity, and clinical visits. As such, caution is warranted when interpreting our results from the pandemic period, as search and health care usage behaviors were likely atypical during this time.
Limitations
Our study has several limitations. First, RSVs may not fully capture the motivations behind individual searches, as users may be influenced by seasonal trends, media coverage, or general health concerns rather than actual symptoms. While social media platforms like YouTube are increasingly used by younger populations for health-related information, our study could not incorporate such data due to limitations in available sources. Furthermore, potential confounding factors such as pandemics or public health campaigns were not controlled for. Nevertheless, previous research supports that individuals often turn to search engines early in the symptom experience, suggesting that, despite these limitations, RSVs remain a valuable proxy for monitoring population-level health trends when interpreted cautiously. Along with that, NAVER is the dominant search engine in South Korea; our findings may not be directly generalizable to other regions where different platforms are more commonly used. However, the underlying methodology—leveraging AI-generated symptom synonyms and analyzing their temporal relationship with health outcomes—can be adapted to platforms. Future cross-platform comparative studies are warranted to examine the consistency and reliability of search behavior across different sociotechnical contexts. In addition, the study does not account for potential misspellings or alternative phrasings, which could have resulted in underestimating actual search volumes. These factors may have affected the comprehensiveness of the data and the overall accuracy of the findings. Second, using ChatGPT carries the risk of providing incorrect information, exhibiting algorithmic bias, and perpetuating biases present in its training data [21,47]. Nevertheless, GPT-4 demonstrated higher response validity than other AI chatbots, with a correct answer rate of 82.2% in medical knowledge, indicating its potential utility in medical education [48,49]. This study used search terms suggested by ChatGPT, rather than relying solely on scientific MeSH (Medical Subject Headings) terms, to include a wide range of potential search terms. ChatGPT uses a large language model and the deep learning architecture developed by OpenAI to search across an immense collection of real-time data and is designed to respond to natural language queries with almost identical words used by general people [22]. To address this issue, we assembled a multidisciplinary team comprising AI data scientists, epidemiologists, and a medical doctor as coauthors for our study, fostering thorough discussions to guarantee a balanced and thorough analysis of the data. Third, the NEDIS database includes data from only 40% of EDs (211 out of 522), primarily large urban hospitals [50]. This coverage may introduce a bias toward urban populations, underrepresenting rural and smaller health care facilities. Patient populations in rural areas tend to be older, with higher rates of chronic conditions, which could result in different health-seeking behaviors and diagnostic patterns compared to urban populations [51]. In addition, rural EDs often rely on general emergency physicians who may operate with few diagnostic resources and less specialized equipment. These disparities could influence the accuracy and completeness of data reported to NEDIS. Consequently, the findings of this study may not fully reflect health care usage trends in rural areas or smaller health care facilities. Fourth, the 4-year gap between the data collection period (2017–2021) and manuscript submission (2024) reflects the rapidly evolving nature of digital health-seeking behaviors. The emergence and widespread use of large language models, such as ChatGPT, may have altered how individuals search for health information on the web. As conversational AI tools become more prevalent, future studies should investigate how these shifts affect search engine usage patterns. While our findings remain relevant for understanding pre-AI behaviors, longitudinal comparisons with post-AI data could enhance digital surveillance approaches. Fifth, the COVID-19 pandemic has profoundly influenced health care usage and public health awareness. Behaviors such as an increased reliance on telemedicine, a heightened focus on infectious disease symptoms, and a greater public familiarity with digital health solutions have altered the generalizability of findings based on pre- and early-pandemic data. However, as the pandemic subsided, public interest in digital health solutions and symptom searching did not sustain at peak levels, potentially impacting the applicability of pre-2024 search patterns to current contexts [44,52].
Despite these limitations, our study provides empirical evidence supporting the utility of ChatGPT-generated search terms in identifying colloquial synonyms that are not captured by MeSH terms. By demonstrating strong correlations between the RSVs of these terms, we highlight ChatGPT’s potential to complement traditional search term generation methodologies in health informatics. This approach addresses the limitations of traditional methods, including Google Trends studies, which often rely on static and predefined keywords. By contrast, our approach uses ChatGPT to dynamically adapt to changes in search behaviors and emerging terms. This not only reduces the reliance on manual efforts but also ensures that the selected terms are reflective of diverse and evolving search behaviors. In addition, we incorporated a validation process using RSV data to ensure the generated terms were both relevant and representative of real-world users’ search behaviors. By demonstrating significant correlations and timely associations between specific diarrhea-related search terms and ED visits, the study validates the potential of AI to enhance symptom surveillance. Moreover, our analyses of the age-based distribution of search engine users and ED visitors provide valuable demographic insights, highlighting which age groups are most likely to seek health information on the web and require medical care. Unlike global platforms such as Google, NAVER reflects local cultural and linguistic nuances, making it particularly well-suited for capturing region-specific search behaviors [53]. In addition, our study offers insights into how internet-based syndromic surveillance may help address both the digital divide and rural underrepresentation in public health data. Rural populations often face limited access to broadband, digital tools, and clinical infrastructure, which can hinder both their web-based health engagement and inclusion in clinical surveillance systems [54]. However, symptom search data from NAVER can serve as a supplementary data stream, particularly in areas where ED data are incomplete or unavailable [55]. In addition, by using ChatGPT to generate colloquial and culturally relevant synonyms, our study enhances the inclusivity of web-based surveillance. Previous research has shown that laypeople, especially those in older or rural populations, often use nonstandard language to describe symptoms [56]. ChatGPT’s ability to capture such expressions increases the sensitivity of surveillance models across linguistic and educational divides.