公共舆论对新冠疫情非新冠类医疗服务利用的影响研究
The impact of public discourse on healthcare utilization during the COVID-19 pandemic
研究将德国新冠公共舆论的语调与关注度关联至非新冠医疗利用,发现二者存在系统性关联。该研究通过对比不同媒体来源及医疗领域异质性,揭示了舆论动态对非新冠诊疗行为的潜在影响机制。
Introduction
The COVID-19 pandemic put exceptional pressure on healthcare systems and disrupted routine medical care far beyond services directly related to COVID-19. Evidence from a range of countries (including Germany) shows substantial declines in the use of healthcare services during the pandemic, especially in its early phases1,2,3,4,5,6,7. Utilization of preventive care such as cancer screening and routine vaccinations declined, and participation in programs for chronic conditions was reduced8,9. These disruptions raise concerns about delayed diagnoses, missed prevention opportunities, and ultimately worse longer-term health outcomes, making it important to understand not only when utilization fell, but also what shaped care-seeking decisions under pandemic conditions and thus how care-seeking behavior can be encouraged.
Most explanations for these declines in healthcare utilization emphasize structural and epidemiological conditions, such as changes in provider capacity10,11, restrictions on mobility12,13, and perceived infection risk in healthcare settings14,15,16,17. At the same time, decisions about seeking care are shaped by the information environment in which people interpret risks and make choices. Throughout the pandemic, people were exposed daily to news reports and social media discussions about case numbers, deaths, hospital pressure, vaccination debates, and policy conflicts18. Such communication does not only inform; it can also shape perceived threat, trust in institutions, and the perceived safety of healthcare settings19,20,21,22,23.
Empirical studies that link public discourse to healthcare utilization beyond COVID-specific care remain rare. While linguistic and communication research has documented how COVID-19 was discussed across government statements, traditional media, and social media, including sentiment trends and emotional framing24,25,26, as well as shifting topic structures and metaphors27, and politicized or conspiratorial narratives28,29, it rarely tests whether discourse dynamics translate into measurable changes in healthcare-seeking behavior. A related but distinct line of research in infodemiology demonstrates that social and traditional media signals can track and sometimes predict disease dynamics, with successful applications to outbreaks such as influenza, measles, and Zika30,31,32,33,34,35,36. These studies establish the predictive value of discourse-based indicators for epidemiological monitoring, but they typically do not connect news coverage and social media discourse to administrative utilization outcomes. Where discourse has been linked more directly to health behavior, the focus has largely been on COVID-specific decisions, especially vaccination and adherence to protective measures, showing associations between vaccine-related online discourse and vaccine attitudes as well as between media exposure and compliance37,38,39,40,41,42,43. By contrast, across these strands of work, population-level links between sustained discourse dynamics and non-COVID healthcare utilization remain underexplored.
We address this gap by linking large-scale measures of public discourse to administrative data on healthcare utilization in Germany during the COVID-19 pandemic. We test whether changes in the emotional tone and topical salience of COVID-19 discourse across newspapers and Twitter were systematically associated with changes in healthcare-seeking behavior, after accounting for epidemiological dynamics, policy conditions, and calendar effects. Beyond establishing these associations, the study makes three key contributions. It (i) compares multiple media sources that differ in agenda-setting potential and audience reach, (ii) evaluates heterogeneity across healthcare domains with different levels of behavioral discretion, and (iii) distinguishes relative attention from evaluative tone to clarify which aspect of discourse aligns with which utilization patterns.
We define public discourse as the shared stream of communication across media channels, including news reporting, commentary, and public discussion, and select from this stream all discourse on and around COVID-19. Public discourse may be linked to healthcare behavior through at least two mechanisms. First, emotional tone, meaning whether discourse is predominantly negative, neutral, or positive, can shape threat perceptions and behavioral responses44,45,46. While coverage emphasizing danger and overwhelmed hospitals can cause individuals to perceive healthcare settings as risky environments and defer services that feel optional47, threat-focused communication can also heighten health anxiety and perceived vulnerability, which may increase protective motivation and health vigilance and thereby prompt care-seeking for some services47,48. Second, topic salience, that is how much attention COVID-19 receives relative to other content, can shift what people perceive as urgent or important49,50. When pandemic coverage dominates the news and social media, other health concerns may recede in perceived priority, leading individuals to postpone preventive and routine care or assume that healthcare systems are focused exclusively on COVID-19, with adverse effects on health outcomes. Heightened salience may, however, also increase demand for related protective behaviors, such as vaccination against other respiratory infections51,52.
These mechanisms imply that healthcare services should differ in their sensitivity to discourse. Preventive services such as cancer screening are often deferrable, whereas chronic disease care is more tightly constrained by medical necessity. Services also differ in contact intensity, such that shared spaces or invasive procedures may feel riskier during a pandemic, and in perceived relevance to the threat, such that respiratory vaccinations may become more salient when infectious disease risks dominate public attention. Sustained COVID coverage may also crowd out attention to non-pandemic health needs and signal system strain, making deferrable services particularly likely to be postponed53,54.
Germany is an instructive setting for examining discourse–utilization relationships because its near-universal health insurance coverage limits financial access barriers, and the federal structure created meaningful regional and temporal variation in pandemic policies and epidemic dynamics that can be addressed using fixed-effects designs55,56. The German media environment further enables comparison across distinct information channels. We distinguish national newspapers, which are plausibly more influential for agenda-setting dynamics, from regional newspapers that may reflect locally salient concerns. Moreover, we analyze the daily tabloid Bild separately, as it is by far the newspaper with the highest circulation and its reporting is widely considered to be influential in German politics and society50,57,58.
We complement newspaper coverage with data from Twitter to capture rapid, user-generated expression, including direct reactions to news events, peer-to-peer discussion, and grassroots concerns that may not surface in edited news content59,60,61,62. While Twitter users are not representative of the general population, the platform provides high-frequency information on public attention and emotional tone that can be aligned with regional healthcare utilization time series. Although Facebook is more widely used in Germany63, it does not provide comparable researcher-accessible historical data and geolocation metadata at scale for reproducible regional time-series analyses64,65. Moreover, while Twitter discourse was generally public, Facebook posts often have restricted public visibility, making Facebook a medium with arguably limited influence on public discourse. Twitter is therefore a superior source of evidence on public discourse, allowing a transparent linkage between discourse measures and regionally resolved utilization data.
We operationalize public discourse using two features: sentiment (negative, neutral, positive) and topic salience (the share of COVID-related content out of all discourse), computed separately for Twitter, national newspapers, regional newspapers, and Bild. We link these measures to administrative utilization data from January 2019 to June 2022 and estimate fixed-effects panel regressions of monthly percentage changes in service use, controlling for epidemic dynamics, policy restrictions, calendar effects, weather anomalies, and month and regional fixed effects. We examine service categories that differ in deferability, contact intensity, and pandemic relevance, including preventive screenings, vaccinations, chronic care, psychotherapy, and general and specialized outpatient care.
This study addresses three research questions. (1) Do sentiment and topic salience predict healthcare utilization beyond epidemiological and structural factors? (2) Do associations differ systematically across service types? (3) Do associations vary by media source?
Based on the mechanisms outlined above, we expect stronger associations for discretionary, contact-intensive services, and weaker associations for medically necessary services, while heightened pandemic salience may be positively associated with uptake of respiratory vaccinations through spillover effects. We further expect heterogeneity across national newspapers, regional newspapers, Bild, and Twitter, while treating the direction and magnitude of source differences as an open empirical question.
Methods
Data sources and corpus construction
We compiled two text corpora covering January 2019 to June 2022, including a full pre-pandemic baseline year and 30 months of the pandemic period. This design allows us to quantify discourse change relative to pre-pandemic patterns and to examine discourse–utilization associations across multiple pandemic waves.
The newspaper corpus was retrieved from LexisNexis and comprised 367,000 German-language articles from 59 outlets. LexisNexis applies automated duplicate and near-duplicate detection and flags repeated and syndicated content. We used these flags in preprocessing and additionally removed exact duplicates in R using standard duplicate detection based on article metadata and text fields to ensure that repeated items were fully excluded. After deduplication, the corpus contained approximately 177 million words. For analysis, newspapers were grouped into national outlets, regional outlets, and Bild.
The Twitter corpus consisted of German-language tweets posted from Germany over the same period that contained location metadata. Tweets were collected via Twitter API v2 using Academic Research access and geolocation filters. We excluded retweets. In practice, retweets are largely absent in geolocation-filtered collections because they typically do not carry usable location metadata, which is required for mapping tweets to regions. We also excluded verified accounts to reduce institutional messaging and focus on user-generated content, and we filtered out unusually high-frequency accounts by excluding accounts that posted more than 10 tweets per day on average. Beyond these account-level filters, we did not apply dedicated bot detection procedures; some automated or coordinated activity may therefore remain in the corpus. Bot detection and removal was routinely performed by Twitter. Any additional attempts to remove potential bot-related activity would have limited effectiveness due to the lack of a ground truth.
Location metadata (geotagged coordinates and user-declared locations, where available) were mapped to German administrative districts (Kreise; NUTS-3) and then assigned to federal states (NUTS-1) using the EU’s Nomenclature of Territorial Units for Statistics (NUTS), enabling alignment with the healthcare utilization data. The final Twitter corpus contained over 19 million tweets and approximately 326 million words.
Identifying COVID-related discourse and measuring topic salience
COVID-related content was identified using a contrastive keyword approach based on month-matched keyness analyses. For each month from January 2020 to June 2022, we treated that month as the target corpus and compared it to the corresponding calendar month in 2019 as the reference corpus, yielding 30 month-level comparisons per platform. Keyness was quantified using log-likelihood (LL) as the inferential criterion for overrepresentation and log ratio (LR) as an effect-size measure66. Candidate keywords were retained if they met three thresholds: LL > 3.84 (p < 0.05), LR > 1 (i.e. at least a doubling in relative frequency), and a minimum frequency of 10 occurrences in the target month67,68. To prioritize robust vocabulary rather than month-specific spikes, we retained only those terms that appeared as significant keywords in at least three of the 30 monthly comparisons.
We conducted the procedure separately for newspapers and Twitter. The resulting keyword sets largely overlapped, while the Twitter-derived list additionally contained abbreviated forms and hashtags (for example, “c19” and platform-specific variants). All retained candidate terms were manually reviewed for COVID relevance. Two coders evaluated items and resolved disagreements through discussion. COVID-relevant keywords were then assigned to semantic fields based on their primary reference. For the analyses reported in this paper, we used the three largest and most clearly COVID-specific fields: COVID-19 names and terminology (e.g., “coronavirus”, “pandemic”, “SARS-CoV-2”), vaccination-related terms (e.g., “vaccine”, “immunization”, “booster”), and non-pharmaceutical interventions (e.g., “lockdown”, “mask”, “quarantine”). The final list contained 616 terms (COVID-19 names: n = 74, non-pharmaceutical interventions: n = 388, vaccination: n = 154). The full keyword lists, semantic-field assignments, and frequency information are reported in Supplementary Data 1.
A text was classified as COVID-related if it contained at least one term from the final keyword set. Topic salience was then operationalized as the proportion of COVID-related content within each source and time period. For Twitter, salience was computed as the share of COVID-related tweets among all geolocated German-language tweets in the same period. For newspapers, salience was computed as the share of COVID-related articles among all articles published in the same period, calculated separately for national outlets, regional outlets, and Bild.
Because newspaper outlets differ substantially in audience reach and because consistent circulation data were not available for all newspapers (and do not capture online readership or regional variation in exposure), we additionally constructed a visibility-weighted salience measure for newspapers using Google Trends. Specifically, we queried each outlet name in Google Trends (Germany, 2020–2022) using a common benchmark query (“Frankreich”) to obtain comparable relative attention indices across outlets. These indices were used to compute, for each period, the share of COVID-related coverage weighted by outlet-level attention (i.e., the proportion of COVID-related articles in a category after weighting each outlet’s contribution by its relative Google Trends attention). This procedure yields a pragmatic proxy for outlet visibility that does not rely on print circulation and is intended to approximate differences in audience reach and online attention. Unweighted salience was retained for descriptive comparisons.
Sentiment annotation and modeling
Sentiment was modeled as a three-class outcome (negative, neutral, positive) to capture the emotional tone of COVID-related discourse. For Twitter, sentiment was classified for each tweet. For newspapers, sentiment was classified at sentence level because emotional tone can vary within articles.
To create training data, we manually annotated 10,000 tweets and 10,000 newspaper sentences sampled proportionally across time periods and sources. Each item was labeled independently by three trained raters using detailed annotation guidelines. Raters classified sentiment based on the overall emotional valence of the text: negative (expressing concern, fear, criticism, or pessimism), neutral (factual reporting or emotionally ambiguous content), or positive (expressing optimism, relief, support, or encouragement). The final label for each item was assigned by majority vote across the three raters. Items that remained ambiguous after annotation were excluded from model training, yielding 8,902 tweets and 8,375 newspaper sentences. In the final training labels, tweets were 34.0% negative (n = 3,021), 57.0% neutral (n = 5,073), and 9.1% positive (n = 808); newspaper sentences were 20.9% negative (n = 1,750), 69.4% neutral (n = 5,814), and 9.7% positive (n = 811). The full annotation guidelines (translated English version) are provided in Supplementary Methods.
Inter-annotator agreement, computed prior to majority voting, was moderate (Fleiss’ κ = 0.47 for tweets; κ = 0.51 for newspaper sentences). This level of agreement is typical for sentiment annotation tasks, particularly in crisis communication contexts where evaluative judgments are inherently subjective and texts often contain mixed or ambiguous emotional cues69,70. Using three raters and majority voting was intended to increase label stability while retaining a simple and interpretable label set suitable for time-series aggregation.
The majority-vote labels were used to train two supervised sentiment classifiers (one for tweets and one for newspaper sentences) based on long short-term memory (LSTM) networks, which model sequential information in text. Texts were lowercased and cleaned (e.g., URLs/HTML removed) and then tokenized and padded. Models were trained on a stratified 70/30 train–holdout split with early stopping and tuned on a validation set; performance was evaluated on held-out data using precision, recall, and F1 by class. Overall performance was strong (weighted F1 = 0.85 for Twitter; weighted F1 = 0.83 for newspapers), indicating reliable capture of sentiment patterns across both corpora.
To reflect differences in reach and engagement across sources, sentiment was weighted prior to aggregation. For newspapers, outlet-level weights were derived from Google Trends as described above. The resulting relative search interest indices were used to weight outlet-level sentiment before aggregating to national, regional, and Bild time series. For Twitter, predicted tweet sentiment was weighted by engagement, operationalized as log-transformed tweet-level retweet_count metadata (log(retweets + 1)). Retweet counts provide a measure of message diffusion and visibility within the platform71. Logarithmic transformation reduces the influence of outliers while preserving the ordinal ranking of engagement.
Healthcare utilization outcomes
Healthcare utilization data were obtained from the Central Research Institute of Ambulatory Health Care in Germany (Zentralinstitut für die kassenärztliche Versorgung), which maintains comprehensive records of outpatient physician visits for the statutory health insurance system. Data cover monthly outpatient visits from January 2019 to June 2022 for all KV regions (Kassenärztliche Vereinigungen, regional associations of statutory health insurance physicians) in Germany except Mecklenburg–Western Pomerania, which was excluded due to incomplete reporting.
We examined five broad categories of healthcare utilization that vary theoretically in urgency, contact intensity, and pandemic relevance: (1) Preventive care: mammography screening, colonoscopy screening, skin cancer screening, and pediatric preventive check-ups (U-examinations). These services are largely discretionary and can typically be postponed without immediate health consequences. (2) Vaccinations: influenza vaccination, pneumococcal vaccination, and other routine vaccinations (excluding COVID-19 vaccines). These preventive services may become more salient during respiratory disease outbreaks. (3) Chronic disease care: disease management program (DMP) participation, hemodialysis, and peritoneal dialysis, and oncological patient care. These services are medically necessary and difficult to defer. (4) Psychotherapy: individual psychotherapy and group psychotherapy. These vary in contact intensity and perceived urgency. (5) General and specialized care: general practitioner (GP) visits, specialist consultations, psychotherapeutic treatment, prenatal care, pediatric visits, and emergency department visits.
To account for changes in population size and to facilitate interpretation, the dependent variable was constructed as the monthly percentage change in utilization relative to the same month in 2019, calculated separately by service type and KV region:
$$\:{\text{Change}}_{r,t}=\frac{{\text{Utilization}}_{r,t}-\:{\text{Utilization}}_{r,t\_2019}}{{\text{Utilization}}_{r,t\_2019}}*\:100$$
where r indexes KV region, t indexes months throughout the pandemic, and t_2019 indexes the corresponding months in the baseline pre-pandemic year of 2019. This specification isolates pandemic-period changes from baseline seasonal patterns and expresses changes in a readily interpretable metric. Population-level counts of insured persons by KV region were obtained from the National Association of Statutory Health Insurance Funds (GKV-Spitzenverband) and used to verify that utilization changes were not driven by shifts in insurance coverage.
Control variables
We included covariates capturing pandemic dynamics, policy conditions, and calendar effects that may confound the relationship between discourse and healthcare utilization.
Data on confirmed COVID-19 infections and deaths were obtained from the Robert Koch Institute (RKI), Germany’s federal public health agency. Case and death counts were aggregated to the federal state level and expressed as weekly rates per 100,000 population. Because care-seeking behavior is likely to respond to published pandemic statistics with a delay, we used one-week-lagged case and death rates. Weekly rates were then aggregated to monthly values to match the temporal resolution of the healthcare utilization outcomes.
Policy conditions were captured using state-level public health regulations and official documentation, drawing on press releases and law and regulation gazettes published on the federal states’ websites. We operationalized policy stringency as whether a mobility restriction was in place. Because mobility restrictions varied widely across states and over time, for example, in whether limits were defined in terms of households, persons, or both, we focused on the most restrictive and comparatively uniform gathering rule: allowing meetings with at most one person from outside the household. For each state and month, we calculated the share of days for which this restriction applied and included this measure as a control variable.
We also controlled for public holidays and school holidays, which affect healthcare utilization and may co-vary with public discourse. Public holidays were obtained from official federal and state calendars maintained by the German Trade Union Confederation (DGB), and school holidays (including summer, autumn, winter, and spring breaks) were obtained from the Standing Conference of Ministers of Education and Cultural Affairs (KMK). Because the dependent variable is defined as the change in utilization relative to the same month in 2019, we constructed holiday controls as changes relative to the corresponding month of 2019. Specifically, for each KV region and month, we computed the change in the share of days that were a public holiday and the change in the share of days that were a school holiday relative to the same calendar month in 2019.
Weather conditions may influence healthcare-seeking behavior, particularly for discretionary services72,73. We included three weather controls (sunshine hours, mean temperature, and precipitation) using data from the German Weather Service (DWD), aggregated to the federal state level. Analogous to the holiday measures, weather variables were constructed as deviations from the corresponding month in 2019 and included to capture unusual conditions beyond normal seasonal patterns.
For comparability of effect sizes across predictors, all continuous covariates (including case and death rates, the contact restriction measure, holiday measures, and weather anomalies) were standardized as z-scores prior to estimation. All models included month fixed effects to capture recurring seasonal patterns and KV-region fixed effects to account for time-invariant regional differences in healthcare infrastructure, provider density, population health, and baseline care-seeking norms.
Statistical analysis
We estimated linear fixed-effects panel regressions predicting the monthly percentage change in healthcare utilization by KV region and service type. The baseline model included control variables and fixed effects. The full model added the discourse measures, including sentiment and topic salience, disaggregated by source (Twitter, national newspapers, regional newspapers, and Bild). Robust standard errors clustered at the KV-region level were used throughout.
As additional specifications, we estimated models including sentiment × salience interaction terms and models excluding weather controls; results are reported in Supplementary Data 2 and were substantively consistent with the main findings.
All data processing and analyses were conducted in R (version 4.4.1) and StataMP 18. Corpus analysis relied on quanteda74, sentiment modeling used keras75 and tensorflow76, and regressions were performed using Stata with diagnostic checks for heteroskedasticity and multicollinearity.
Results
Patterns in public discourse across media sources
Before linking discourse to healthcare behavior, we characterize how COVID-19 was discussed across newspapers and Twitter during the pandemic. We first focus on monthly topic salience, defined as the proportion of COVID-related content within each source and month (see Fig. 1). COVID-related content rose sharply across all sources at pandemic onset. In national and regional newspapers, COVID salience peaked in spring 2020 at approximately 65–75% of all articles, remained elevated through 2021, and declined gradually into mid-2022. Bild showed a similar trajectory but consistently lower salience (peaking around 60%), reflecting its broader content mix. Twitter displayed substantially lower COVID salience throughout, rarely exceeding 10% even during peak periods, indicating that pandemic discussion constituted a smaller fraction of overall public conversation on the platform.
COVID-19 topic salience across media sources over time. Monthly share of COVID-related content within each source from January 2020 to June 2022. Lines show topic salience for Twitter (red), Bild (green), national press (blue), and regional press (purple).
Moving on to sentiment in COVID-related discourse, the data indicate that, across all sources, mean sentiment remained on the negative side of the scale for most months, while sources differed in level and variability (see Fig. 2). Twitter was consistently more negative (mean ≈ -0.25 to -0.35) than newspapers. National newspapers, regional newspapers, and Bild clustered closer together with moderate negativity (mean ≈ -0.07 to -0.25) and relatively stable month-to-month patterns. Twitter showed greater variability, with sharper fluctuations corresponding to specific events or policy announcements.
Mean sentiment in COVID-related discourse across media sources over time. Monthly average sentiment scores (− 1 = negative to + 1 = positive) from January 2020 to June 2022. Lines show Twitter (tweet-level; red), Bild (green), national press (blue), and regional press (purple). The dashed horizontal line marks neutral sentiment (0).
To illustrate these aggregate trends, we provide examples for each sentiment class from both corpora (English translations; identifiers removed for Twitter). Positive items commonly pointed to small markers of continuity and regained routines under ongoing restrictions:
-
1.
All participants were thrilled that the [film festival] could finally take place despite all COVID-related restrictions. (Nürnberger Zeitung, 2021-07-20)
-
2.
Following all COVID rules, we’re back at the office. It’s nice to see everyone again. (Twitter; identifiers removed).
Neutral items largely consisted of descriptive reporting and practical guidance without explicit stance-taking:
-
(3)
Even after the end of the COVID crisis, the German economy expects significantly fewer business trips than before. (Die Welt, 2021-05-31)
-
(4)
I was told everything was scanned the first time, so for the second appointment you only need the confirmation and your vaccination card. (Twitter; identifiers removed)
Negative items most often emphasized risk, strain, or dissatisfaction with pandemic management, either through institutional concern about system capacity or through direct critique and harm-focused projections.
-
(5)
There is great concern that, even if not SARS-CoV-2 alone, the combined burden of influenza and COVID-19 could bring the health system to its knees. (Der Spiegel, 2020-09-26)
-
(6)
This graphic unfortunately shows that the virus is moving faster than our politics. While officials keep promoting a strategy focused on older risk groups, more and more children are being infected with uncertain long-term consequences. (Twitter; identifiers removed).
These examples clarify how the sentiment labels map onto recurring discourse patterns in the data and help contextualize the more moderate, stable newspaper tone versus the sharper, more variable evaluations observed on Twitter.
Together, these descriptive patterns motivate source-specific analyses: newspapers show high and sustained COVID salience with comparatively stable, moderately negative sentiment, whereas Twitter exhibits lower salience but more negative and volatile tone. Bild is less COVID-focused than quality press but follows similar sentiment trajectories, consistent with its broader content mix.
Healthcare utilization during the pandemic.
Healthcare utilization declined most sharply during the first pandemic wave, with the steepest drop (− 24% relative to the 2019 baseline) in April 2020 (see Fig. 3). This early drop occurred before COVID-19 death rates peaked, consistent with a strong initial disruption of routine care at the onset of the pandemic. Utilization then recovered quickly and fluctuated around baseline levels through late 2020. A second pronounced downturn occurred in early 2021 (January: −15%; February: −12%), coinciding with the winter 2020–21 peak in COVID-19 mortality. Thereafter, utilization rebounded substantially in mid-2021 (June: +21%) and remained often at or above baseline from late 2021 into mid-2022, despite renewed increases in mortality in late 2021; occasional negative deviations persisted (e.g., April 2022: −7%).
Healthcare utilization and COVID-19 mortality over time. Monthly percentage difference in total outpatient visits relative to the 2019 baseline (blue line; left y-axis) and weekly COVID-19 death rates per 100,000 population (gray line; right y-axis) in Germany from January 2020 to June 2022. The horizontal black line indicates no change from baseline (0%).
Healthcare utilization varied markedly between services (see Fig. 4, relative to the 2019 baseline). The largest reductions were observed for emergency visits (−19%), DMP trainings (−12%), and prenatal care (−11%), while pediatric visits declined more moderately (−3%). In contrast, several preventive services showed increases, including skin cancer screening (+5%), colonoscopy screening (+3%), and child preventive check-ups (+2%). Vaccination uptake also increased, particularly for pneumococcal (+14%) and influenza vaccinations (+4%). Among therapy services, group therapy increased by approximately 10%, while individual and psychotherapeutic treatment increased by around 3%. Hemodialysis showed only a small increase (+1%), whereas peritoneal dialysis increased by approximately 5%.
Overall change in healthcare utilization by service type during the pandemic. Overall percentage change in utilization for each healthcare service from 2020 to mid-2022 relative to the 2019 baseline. Bars to the right of zero indicate higher utilization than baseline (blue) and bars to the left indicate lower utilization (red). Services are grouped by domain (preventive care, chronic care, vaccinations, general and specialized care, and therapy); the dashed vertical line marks zero change.
These patterns confirm substantial heterogeneity across healthcare domains and provide a basis for testing whether discourse–utilization associations differ systematically by service type.
Baseline model: epidemiological and structural associations.
We first estimated a baseline fixed-effects model including epidemiological indicators (one-week-lagged COVID cases and death rates), policy measures (contact restrictions), calendar effects (public and school holidays), and weather anomalies, along with month and region fixed effects. This baseline establishes how much of the variation in utilization can be accounted for by epidemiological dynamics and structural conditions alone, providing the benchmark for assessing the added contribution of discourse predictors.
In a model for aggregate healthcare utilization, the one-week-lagged case rate was positively associated with utilization (β = 3.26, SE = 0.47, p < 0.001), likely reflecting increased testing and COVID-related consultations (see coefficient estimates in Fig. 5). In contrast, the lagged death rate was strongly negatively associated with utilization (β = -3.27, SE = 0.67, p < 0.001), indicating that care-seeking declined during periods of higher mortality.
Baseline predictors of aggregate healthcare utilization. Coefficient estimates (points) with 95% confidence intervals (lines) from the baseline model predicting monthly percentage change in total outpatient utilization relative to the 2019 baseline. Predictors include lagged COVID-19 case and death rates, mobility restrictions, public and school holiday indicators, and weather deviations. Red markers indicate statistically significant associations (p < 0.05) and gray markers indicate non-significant associations; the dashed vertical line marks a coefficient of zero.
Calendar effects were pronounced: public holidays (β = -1.53, SE = 0.56, p < 0.01) and school holidays (β = -3.01, SE = 0.30, p < 0.001) both reduced utilization, consistent with reduced provider availability and altered scheduling behavior. The mobility restriction indicator was negative but not statistically significant (β = -0.97, SE = 0.72, p = 0.18), suggesting limited additional explanatory power once epidemiological dynamics and calendar patterns were controlled. Among weather variables, only sunshine deviation showed a significant negative association (β = -2.77, SE = 0.68, p < 0.001); temperature (β = -0.57, SE = 0.55, p = 0.30) and precipitation deviations (β = -0.17, SE = 0.75, p = 0.83) were not significant. Overall, the baseline model explained a substantial share of variation in aggregate utilization (adjusted R² = 0.68), establishing a strong benchmark for assessing discourse contributions.
Adding discourse: model fit improvements.
We extended the baseline model by adding sentiment and topic salience measures for each media source (Twitter, national newspapers, regional newspapers, Bild). Additionally, baseline models and extended models were computed for 19 individual healthcare services across five categories. The added explanatory power of the discourse variables in the extended models is derived as the difference in adjusted R² values between baseline and full models (see Fig. 6).
Model fit improvement after adding discourse predictors. Adjusted R² values from baseline models (gray ×) and discourse-extended models (red diamonds) for each utilization outcome. Outcomes are grouped by service domain (all services, preventive care, chronic care, vaccinations, general and specialized care, therapy); vertical dashed lines separate domains.
For aggregate utilization, adjusted R² increased from 0.68 to 0.74 (+ 0.06). Among individual healthcare services, the largest gains appeared for discretionary services requiring active patient decisions: mammography screening (+ 0.27, from 0.35 to 0.62), colonoscopy screening (+ 0.07, from 0.65 to 0.72), and skin cancer screening (+ 0.10, from 0.67 to 0.77). Models for vaccination outcomes also improved substantially: influenza vaccination (+ 0.08, from 0.32 to 0.40), pneumococcal vaccination (+ 0.12, from 0.30 to 0.42), and other vaccinations (+ 0.02, from 0.46 to 0.48). Models for psychotherapy services also showed notable improvements: group therapy (+ 0.13, from 0.64 to 0.77) and individual therapy (+ 0.05, from 0.63 to 0.68), indicating that psychotherapy utilization is more strongly associated with discourse measures than most chronic care services. By contrast, models for medically necessary services showed minimal gains: hemodialysis (+ 0.02, from 0.75 to 0.77), peritoneal dialysis (+ 0.01, from 0.71 to 0.72), and DMP trainings (+ 0.02, from 0.75 to 0.77), indicating that discourse has limited predictive value where clinical necessity dominates. Overall, discourse measures added explanatory power beyond structural predictors, with the largest gains for discretionary or deferrable services and the smallest gains for medically necessary care.
Service-specific discourse associations.
We next examine whether these associations vary systematically by service type and media source (see Figs. 7, 8 and 9 for coefficient estimates and Supplementary Data 2 for a comprehensive summary). We focus on statistically significant associations (p < 0.05). Coefficients reflect the expected change in utilization (percentage points relative to 2019 baseline) associated with a one standard deviation increase in the discourse measure.
Discourse predictors of preventive care utilization across media sources. Coefficient estimates (points) with 95% confidence intervals (lines) from fixed-effects regressions predicting monthly percentage change in preventive service use relative to the 2019 baseline. Panels show mammography screening, skin cancer screening, colonoscopy screening, and child preventive check-ups. Predictors include sentiment and COVID-19 topic salience measured separately for Twitter, national press, regional press, and Bild. Red estimates indicate statistically significant associations (p < 0.05) and gray estimates indicate non-significant associations; the dashed vertical line marks a coefficient of zero.
Discourse predictors of vaccination uptake across media sources. Coefficient estimates (points) with 95% confidence intervals (lines) from fixed-effects regressions predicting monthly percentage change in influenza vaccinations, pneumococcal vaccinations, and other vaccinations relative to the 2019 baseline. Predictors include sentiment and COVID-19 topic salience measured separately for Twitter, national press, regional press, and Bild. Red estimates indicate statistically significant associations (p < 0.05) and gray estimates indicate non-significant associations; the dashed vertical line marks a coefficient of zero.
Across preventive services (Fig. 7), topic salience measures showed consistent associations for Twitter. A higher relative share of COVID-related discourse on Twitter was negatively associated with utilization for mammography (β = −18.32, p < 0.001), skin cancer screening (β = −2.95, p = 0.001), colonoscopy screening (β = −3.04, p < 0.001), and child preventive check-ups (β = −1.63, p < 0.05). Associations for salience in other sources were more outcome-specific. For child preventive check-ups, regional press salience (β = 0.63, p < 0.05) and Bild salience (β = 0.82, p < 0.05) were positively associated with utilization. For adult screenings, Bild salience was negatively associated with skin cancer screening (β = −4.35, p < 0.001) and colonoscopy screening (β = −2.10, p < 0.001), while mammography showed a positive association with Bild salience (β = 13.61, p < 0.001).
Significant association of sentiment with preventive care were concentrated in national newspapers and Bild. More positive sentiment in Bild was associated with lower utilization across all four preventive outcomes (mammography: β = −9.47, p < 0.001; skin cancer screening: β = −3.05, p < 0.001; colonoscopy screening: β = −3.69, p < 0.001; child preventive check-ups: β = −2.57, p < 0.001). National press sentiment was also negatively associated with mammography (β = −13.20, p < 0.001) and colonoscopy screening (β = −2.02, p = 0.002). Two additional associations differed by source: skin cancer screening showed a positive association with regional press sentiment (β = 0.75, p < 0.1), and a negative association with Twitter sentiment (β = −1.80, p < 0.001). Overall, preventive services were most consistently associated with higher COVID salience on Twitter (negative) and with more positive sentiment in national newspapers and Bild (negative), while salience in newspapers showed more outcome-specific patterns.
Therapy outcomes showed several significant associations across both salience and sentiment (see Fig. 8: note the difference in scale relative to Fig. 7). For group therapy, a higher relative share of COVID-related discourse on Bild (β = −11.52, p < 0.001) and on Twitter (β = −3.19, p < 0.01) was negatively associated with utilization. Sentiment measures were also significant: more positive Twitter sentiment (β = −3.57, p < 0.001) and more positive national press sentiment (β = −2.73, p < 0.01) were associated with lower group therapy utilization.
Discourse predictors of psychotherapy utilization across media sources. Coefficient estimates (points) with 95% confidence intervals (lines) from fixed-effects regressions predicting monthly percentage change in group therapy and individual therapy utilization relative to the 2019 baseline. Predictors include sentiment and COVID-19 topic salience measured separately for Twitter, national press, regional press, and Bild. Red estimates indicate statistically significant associations (p < 0.05) and gray estimates indicate non-significant associations; the dashed vertical line marks a coefficient of zero.
For individual therapy, salience effects differed by source. A higher relative share of COVID-related discourse in Bild (β = −1.02, p < 0.001) and in regional newspapers (β = −0.79, p < 0.001) was negatively associated with utilization, whereas national newspaper salience was positively associated (β = 0.47, p < 0.1). Sentiment effects also varied across sources: more positive Bild sentiment (β = −0.89, p = 0.009) and more positive Twitter sentiment (β = −0.40, p < 0.1) were associated with lower utilization, while regional press sentiment was positively associated (β = 0.36, p < 0.1). Psychotherapy outcomes showed sensitivity to both salience and sentiment, with the strongest negative associations concentrated in high-visibility sources for group settings, whereas individual therapy exhibited more heterogeneous newspaper patterns.
Vaccination outcomes (Fig. 9) exhibited a distinct pattern compared to preventive and therapy services. For influenza vaccinations, a higher relative share of COVID-related discourse on Twitter (β = 56.66, p < 0.001) and in Bild (β = 42.80, p < 0.001) was positively associated with uptake. Twitter sentiment was also positive (β = 22.62, p < 0.01). Pneumococcal vaccinations showed a similar pattern: Twitter salience (β = 27.33, p < 0.001) and Bild salience (β = 25.08, p < 0.001) were positive, and Twitter sentiment was positive (β = 15.39, p < 0.1).
For other vaccinations, the sign pattern differed across sources. Bild salience was positively associated with uptake (β = 1.92, p < 0.01), whereas Twitter salience was negatively associated (β = −4.55, p = 0.001). Twitter sentiment was also negative (β = −1.38, p < 0.1). In contrast to screenings and psychotherapy, vaccination uptake tended to increase when COVID-related attention was higher, particularly on Twitter and in Bild, consistent with spillover toward respiratory protection behaviors.
Discussion
This study asked whether sentiment and topic salience in public discourse around COVID-19 on social and traditional media were systematically associated with healthcare utilization in Germany, beyond epidemiological dynamics, policy conditions, and calendar effects, and whether these associations differed across service types and media sources. Results show that adding discourse measures improved model fit for several outcomes, especially those that require active patient decisions and can be postponed, indicating that the information environment captured additional variance beyond cases, deaths, restrictions, holidays, and weather. Second, associations varied strongly by service domain: utilization of preventive screenings and psychotherapy generally fell when COVID discourse became more salient and, for several outcomes, when coverage shifted toward a less negative (i.e., more neutral/positive) tone, whereas respiratory vaccinations showed the opposite pattern. Third, effects were source-specific: sentiment associations were clearest for newspapers (especially national outlets and Bild) in relation to preventive care, while salience associations were most consistent for Twitter and Bild and were directionally dependent on the service type.
A key implication of these patterns is that discourse appears most relevant where patients retain discretion over whether and when to seek care. This aligns with broader evidence that many people delayed or avoided medical care during the pandemic due to COVID-related concerns, even outside contexts of strict access constraints14,15,16,17. In our models, discourse measures mattered least for medically necessary services such as dialysis, which are difficult to defer, but they substantially improved prediction for screenings and psychotherapy, where perceived risk, uncertainty, and shifting priorities plausibly shape behavior. In this sense, the findings extend previous health communication research22,45,46 by showing that sustained, population-level discourse dynamics correlate with utilization patterns across multiple healthcare domains, not only with explicitly COVID-related behaviors.
The preventive care results point to two distinct pathways for the influence of discourse on HCU (see Table 1 for a summary). Higher COVID salience on Twitter was consistently associated with reduced utilization across all four preventive outcomes, and more positive sentiment in national newspapers and Bild was also linked to lower utilization for several screenings. These measures capture different aspects of the information environment. Salience indicates how dominant the pandemic became in daily communication, consistent with agenda-setting accounts of issue prominence shaping perceived priorities49,50, while sentiment captures whether that communication emphasized threat, loss, criticism, or pessimism (negative) versus factual description (neutral) or reassurance, improvement, and encouragement (positive), in line with work on message framing and evaluative tone shaping health perceptions and responses45,46. Because mean sentiment remained negative for most months, increases in sentiment typically reflect a shift toward less negative coverage rather than clearly positive discourse. The pattern therefore suggests that both sustained COVID attention and shifts in evaluative tone were associated with lower screening uptake. One interpretation is that persistent COVID saturation may crowd out attention to non-pandemic health needs, while periods of less negative coverage may coincide with normalization and reduced urgency to catch up on deferrable preventive services. At the same time, the heterogeneous signs across sources and outcomes indicate that tone likely captures multiple, partially competing processes, underscoring the need for more fine-grained content measures45,46,51.
Psychotherapy showed a similarly discourse-sensitive pattern but with stronger differences between group and individual formats. Group therapy utilization decreased when COVID discourse was more salient in high-visibility sources and when sentiment shifted toward a less negative tone on Twitter and in national newspapers. Individual therapy shared some of these associations but also exhibited more heterogeneous newspaper patterns, including positive associations for some newspaper measures. One plausible interpretation is that group settings are particularly sensitive to perceived infection risk and situational threat, while individual therapy may be buffered by differences in delivery modes and by shifts in perceived need during crisis periods. In our annotation scheme, negative items typically emphasized worsening case trajectories, mortality, hospital burden, policy conflict, and uncertainty, whereas positive items more often highlighted stabilization, effective countermeasures, vaccine availability, and recovery, consistent with prior descriptions of how COVID-19 discourse evolved across social media and news-related communication channels24,25,26,27. Prior work suggests that sustained exposure to COVID-related news and social media can increase psychological strain and anxiety, which provides a plausible bridge between evaluative tone and willingness to enter care settings, particularly for services that can be delayed14,20,21. In addition, COVID communication often included politicized and conspiratorial narratives and recurring themes of institutional distrust, which may have further increased uncertainty about whom to trust and whether healthcare institutions were safe or functioning normally21,28,38.
The vaccination results show a striking divergence from screenings and therapy. Higher COVID salience on Twitter and in Bild was positively associated with influenza and pneumococcal vaccination uptake, and more positive Twitter sentiment was also linked to higher uptake. This pattern is consistent with a spillover account: when attention to infectious disease risk dominates communication, the perceived value of respiratory protection behaviors rises, even when those behaviors target pathogens other than SARS-CoV-251,52. The source differences are informative here. Twitter’s salience and sentiment are likely to capture rapid shifts in grassroots concern and peer-to-peer discussion60,61, which can amplify perceived vulnerability and the salience of preventive actions. Prior research on COVID-19 vaccine discourse on Twitter shows that sentiment, topical focus, and the circulation of persuasive or misleading content vary systematically across networked communities and are linked to vaccine-related attitudes and behavioral outcomes38,39,40,42,43. In parallel, the positive Bild salience associations for respiratory vaccinations and selected preventive services suggest that high-reach tabloid coverage may increase visibility and urgency for particular health behaviors22,23, even when its tone is often emotionally charged.
This helps clarify why effects associated with coverage of the tabloid Bild appeared heterogeneous across domains. Bild showed strong salience associations that were negative for several screenings and psychotherapy outcomes but positive for respiratory vaccinations, suggesting that visibility alone may mobilize action under some conditions while coinciding with reduced discretionary care under others. In addition, the negative coefficients for Bild sentiment indicate lower utilization when COVID-related coverage in Bild became less negative, which suggests that tone may act as a marker of broader pandemic phases (e.g., stabilization, policy transitions, or shifting public attention) rather than mapping straightforwardly onto deterrence through threat alone. One explanation may be that tabloid-style communication can heighten perceived personal relevance and urgency through simplified narratives and prominent cues, potentially increasing responsiveness when coverage coincides with public campaigns, seasonal vaccination timing, or moments when protective behavior is framed as immediately relevant22,45,46,50. This interpretation is also compatible with survey-based research showing heterogeneity in how different media diets relate to compliance, vaccine trust, and pandemic-related behaviors, with quality news, tabloids, and social media channels contributing differently across audiences37,40,41.
Several limitations qualify these conclusions and point to directions for future work. First, the analysis is observational and based on regional aggregates, so causal inference is limited and individual mechanisms cannot be directly tested. Strong controls and fixed effects help limit the relevance of confounding variables but cannot eliminate them; unobserved variables could simultaneously affect discourse and utilization. Second, exposure is approximated rather than measured. Google Trends weights and retweet-based weighting improve the modelling of exposure, but they do not directly observe who consumed which content. Moreover, our Twitter corpus is restricted to geolocated German-language tweets, which likely overrepresents certain user groups and communication contexts; the Twitter results should therefore be interpreted as reflecting a selective subset of platform discourse rather than as representative measures of public opinion. Future studies could improve exposure measurement by combining individual-level media diaries or passive tracking with healthcare utilization records, allowing direct tests of whether media consumption mediates utilization changes. Third, provider-side adaptation is only indirectly captured. Changes in appointment availability, deferred invitations for screening programs, shifts to telehealth, and local practice closures could covary with pandemic attention and could partially explain utilization changes10,77,78,79. Linking claims data to provider capacity indicators, telehealth uptake, and scheduling constraints would help disentangle demand-side responses from supply-side limitations. Finally, our discourse indicators capture broad shifts in how much attention COVID received and how negatively or positively it was discussed, but they do not distinguish what these evaluative signals are directed at. In particular, negative tone can reflect qualitatively different contents, including perceived disease threat, hospital burden and system strain, policy conflict, or institutional distrust, which may plausibly relate to utilization through different pathways. This limitation is especially relevant given that tone–utilization associations were not uniformly consistent with a simple deterrence account. A next step is therefore a more fine-grained content analysis that distinguishes targets and frames of evaluation (e.g., disease threat vs. healthcare-system capacity vs. policy performance) and links these dimensions to utilization outcomes, ideally employing quasi-experimental variation such as sudden policy announcements or differential regional media penetration as sources of identification.
Conclusion
This study connected sentiment and topic salience in COVID-19 discourse across social media (Twitter), as well as national and regional newspapers, to healthcare utilization in Germany from January 2019 to June 2022. Discourse measures predicted utilization beyond epidemiological dynamics, policy conditions, calendar effects, weather anomalies, and fixed effects, with the strongest associations and statistical model improvements occurring for services with greater behavioral discretion. Preventive screenings and psychotherapy were generally used more rarely when COVID discourse was more salient and when tone was more positive, whereas influenza and pneumococcal vaccination uptake increased sharply with higher COVID salience on Twitter and in Bild, and with more positive Twitter sentiment. The results also demonstrate substantial heterogeneity between media types, indicating that where pandemic discourse becomes prominent matters for how it relates to healthcare behavior. Overall, the findings show that public discourse is systematically associated with healthcare engagement during a prolonged emergency and that its associations differ across care domains in ways consistent with deferability, contact intensity, and perceived relevance to infectious disease threat.
Data availability
The Twitter data analyzed in this study cannot be shared publicly due to the end-user agreement of the data provider. Access to the LexisNexis newspaper corpus is restricted and subject to licensing agreements with the provider. The code and aggregated data underlying the findings of this study are available from the corresponding author upon request.
References
Moynihan, R. et al. Impact of COVID-19 pandemic on utilisation of healthcare services: a systematic review. BMJ Open. 11, e045343 (2021).
Nguyen, J. L. et al. Pandemic-related declines in hospitalization for non-COVID-19-related illness in the United States from January through July 2020. PLoS One. 17, e0262347 (2022).
Mattingly, A. S. et al. Trends in US Surgical Procedures and Health Care System Response to Policies Curtailing Elective Surgical Operations During the COVID-19 Pandemic. JAMA Netw. Open. 4, e2138038 (2021).
Yamaguchi, S. et al. Impact of COVID-19 pandemic on healthcare service use for non-COVID-19 patients in Japan: retrospective cohort study. BMJ Open. 12, e060390 (2022).
Kondilis, E., Tarantilis, F. & Benos, A. Essential public healthcare services utilization and excess non-COVID-19 mortality in Greece. Public. Health. 198, 85–88 (2021).
Arsenault, C. et al. COVID-19 and resilience of healthcare systems in ten countries. Nat. Med. 28, 1314–1324 (2022).
Bayindir, E. E. & Schreyögg, J. C. D. Universal Health Coverage, and Ambulatory Care in 2020. Med. Care 60, 413–414 (2022).
Platen, M., Bohlken, J., Hoffmann, W., Kostev, K. & Michalowsky, B. The long-term impact of the COVID-19 pandemic on primary and specialized care provision and disease recognition in Germany. Front Public. Health 10, (2022).
Michalowsky, B., Hoffmann, W., Bohlken, J. & Kostev, K. Effect of the COVID-19 lockdown on disease recognition and utilisation of healthcare services in the older population in Germany: a cross-sectional study. Age Ageing. 50, 317–325 (2021).
Whaley, C. M. et al. Changes in Health Services Use Among Commercially Insured US Populations During the COVID-19 Pandemic. JAMA Netw. Open. 3, e2024984 (2020).
McCabe, R. et al. Adapting hospital capacity to meet changing demands during the COVID-19 pandemic. BMC Med. 18, 329 (2020).
Cantor, J., Sood, N., Bravata, D. M., Pera, M. & Whaley, C. The impact of the COVID-19 pandemic and policy response on health care utilization: Evidence from county-level medical claims and cellphone data. J. Health Econ. 82, 102581 (2022).
Gupta, S., Simon, K. & Wing, C. Mandated and Voluntary Social Distancing during the COVID-19 Epidemic. Brookings Pap Econ. Act. 2020, 269–326 (2020).
Sahakyan, S., Muradyan, D., Giloyan, A. & Harutyunyan, T. Factors associated with delay or avoidance of medical care during the COVID-19 pandemic in Armenia: results from a nationwide survey. BMC Health Serv. Res. 24, 20 (2024).
Zhong, S., Huisingh-Scheetz, M. & Huang, E. S. Delayed medical care and its perceived health impact among US older adults during the COVID ‐19 pandemic. J. Am. Geriatr. Soc. 70, 1620–1628 (2022).
Johnson, K. J. et al. Assessment of the impact of the COVID-19 pandemic on health services use. Public. Health Pract. 3, 100254 (2022).
Gertz, A. H., Pollack, C. C., Schultheiss, M. D. & Brownstein, J. S. Delayed medical care and underlying health in the United States during the COVID-19 pandemic: A cross-sectional study. Prev. Med. Rep. 28, 101882 (2022).
Abbas, A. H. Politicizing COVID-19 Vaccines in the Press: A Critical Discourse Analysis. Int. J. Semiot. Law. 35, 1167–1185 (2022).
Kisa, S. & Kisa, A. A. Comprehensive Analysis of COVID-19 Misinformation, Public Health Impacts, and Communication Strategies: Scoping Review. J. Med. Internet Res. 26, e56931 (2024).
Abbas, J., Wang, D., Su, Z. & Ziapour, A. The Role of Social Media in the Advent of COVID-19 Pandemic: Crisis Management, Mental Health Challenges and Implications. Risk Manag Healthc. Policy. 14, 1917–1932 (2021).
Zhou, Y. et al. Protecting public’s wellbeing against COVID-19 infodemic: The role of trust in information sources and rapid dissemination and transparency of information over time. Front Public. Health 11, (2023).
Wakefield, M. A., Loken, B. & Hornik, R. C. Use of mass media campaigns to change health behaviour. Lancet 376, 1261–1271 (2010).
Wiedicke, A., Stehr, P. & Rossmann, C. Portrayal of risk information and its impact on audiences’ risk perception during the Covid-19 pandemic: A multi‐method approach. Risk Anal. 45, 1545–1557 (2025).
Wicke, P. & Bolognesi, M. M. Covid-19 Discourse on Twitter: How the Topics, Sentiments, Subjectivity, and Figurative Frames Changed Over Time. Front. Commun. (Lausanne). 6, (2021).
Lwin, M. O. et al. Global Sentiments Surrounding the COVID-19 Pandemic on Twitter: Analysis of Twitter Trends. JMIR Public. Health Surveill. 6, e19447 (2020).
Figueiredo, S. Topic modelling and sentiment analysis during COVID-19 revealed emotions changes for public health. Sci. Rep. 14, 24954 (2024).
Schweinberger, M., Haugh, M. & Hames, S. Analysing discourse around COVID-19 in the Australian Twittersphere: A real-time corpus-based analysis. Big Data Soc 8, (2021).
McGlashan, M. et al. COVID-19 vaccine conspiracy theories, discourses of liberty, and the new normal on social media. Linguistics Vanguard. https://doi.org/10.1515/lingvan-2024-0121 (2025).
Orts, M. Á., Vargas-Sierra, C. & Warning or Manipulating in Pandemic Times? A Critical and Contrastive Analysis of Official Discourse Through the English and Spanish News. Int. J. Semiot. Law. 35, 903–935 (2022).
Santillana, M. et al. Combining Search, Social Media, and Traditional Data Sources to Improve Influenza Surveillance. PLoS Comput. Biol. 11, e1004513 (2015).
Pruss, D. et al. Zika discourse in the Americas: A multilingual topic analysis of Twitter. PLoS One. 14, e0216922 (2019).
Paul, M. J., Dredze, M. & Broniatowski, D. Twitter Improves Influenza Forecasting. PLoS Curr. https://doi.org/10.1371/currents.outbreaks.90b9ed0f59bae4ccaa683a39865d9117 (2014).
Mollema, L. et al. Disease Detection or Public Opinion Reflection? Content Analysis of Tweets, Other Social Media, and Online Newspapers During the Measles Outbreak in the Netherlands in 2013. J. Med. Internet Res. 17, e128 (2015).
Ji, X., Chun, S. A. & Geller, J. Monitoring Public Health Concerns Using Twitter Sentiment Classifications. in IEEE International Conference on Healthcare Informatics 335–344 (IEEE, 2013). https://doi.org/10.1109/ICHI.2013.47
Jürgens, P., Jungherr, A. & Schoen, H. Small worlds with a difference. New gatekeepers and the filtering of political information on Twitter. in Proceedings of the 3rd International Web Science Conference 1–5ACM, New York, NY, USA, (2011). https://doi.org/10.1145/2527031.2527034
Culotta, A. Towards detecting influenza epidemics by analyzing Twitter messages. in Proceedings of the First Workshop on Social Media Analytics 115–122 (ACM, 2010). https://doi.org/10.1145/1964858.1964874
Frissen, T., De Coninck, D., Matthys, K. & d’Haenens, L. Longitudinal Evidence of How Media Audiences Differ in Public Health Perceptions and Behaviors During a Global Pandemic. Front Public. Health 8, (2020).
Allington, D., McAndrew, S., Moxham-Hall, V. & Duffy, B. Coronavirus conspiracy suspicions, general vaccine attitudes, trust and coronavirus information source as predictors of vaccine hesitancy among UK residents during the COVID-19 pandemic. Psychol. Med. 53, 236–247 (2023).
Puri, N., Coomes, E. A., Haghbayan, H. & Gunaratne, K. Social media and vaccine hesitancy: new updates for the era of COVID-19 and globalized infectious diseases. Hum. Vaccin Immunother. 16, 2586–2593 (2020).
Green, J. et al. Media use and vaccine resistance. PNAS Nexus 2, (2023).
Ruiz, J. B. & Bell, R. A. Predictors of intention to vaccinate against COVID-19: Results of a nationwide survey. Vaccine 39, 1080–1086 (2021).
Scannell, D. et al. COVID-19 Vaccine Discourse on Twitter: A Content Analysis of Persuasion Techniques, Sentiment and Mis/Disinformation. J. Health Commun. 26, 443–459 (2021).
Hwang, J. et al. Vaccine discourse during the onset of the COVID-19 pandemic: Topical structure and source patterns informing efforts to combat vaccine hesitancy. PLoS One. 17, e0271394 (2022).
Sassenrath, C., Greving, H. & Sassenberg, K. Are you Concerned? Patient Testimonials in Medical Communication Affect Healthy Recipients’ Emotions and Memory. Front Commun. (Lausanne) 3, (2018).
Rothman, A. J. & Salovey, P. Shaping perceptions to motivate healthy behavior: The role of message framing. Psychol. Bull. 121, 3–19 (1997).
Gallagher, K. M. & Updegraff, J. A. Health Message Framing Effects on Attitudes, Intentions, and Behavior: A Meta-analytic Review. Ann. Behav. Med. 43, 101–116 (2012).
Simonovic, N. & Taber, J. M. Psychological Impact of Ambiguous Health Messages about COVID-19. J. Behav. Med. 45, 159–171 (2022).
Rogers, R. W. A Protection Motivation Theory of Fear Appeals and Attitude Change1. J. Psychol. 91, 93–114 (1975).
Asmundson, G. J. G. & Taylor, S. How health anxiety influences responses to viral outbreaks like COVID-19: What all decision-makers, health authorities, and health care professionals need to know. J. Anxiety Disord. 71, 102211 (2020).
Won, J. Y., Kim, H. J., Jeon, E. G., Kim, Y. T. & Lee, Y. R. Media coverage volume and expert risk perception in risk assessment: the moderating role of domain expertise. J. Risk Res. 28, 747–767 (2025).
Mccombs, M. E. & Shaw, D. L. The Agenda-Setting Function of Mass Media. SPublic. Opin. Quarterly 36: 176-187. (1972).
Kromer, C. et al. Impact of COVID-19 on Influenza and Pneumococcal Vaccination of Psoriatic Patients in Germany: Results from Vac-Pso. Vaccines (Basel). 12, 614 (2024).
Ou, M. & Ho, S. S. Conspiracy Beliefs Secondary Risk Perceptions, and Conditional Acceptance of Dengue Vaccine: A Multigroup Comparison Based on Prior COVID-19 Vaccination Experiences. Health Commun. 41, 1–13. https://doi.org/10.1080/10410236.2025.2539462 (2025).
Atherly, A., Van Den Broek-Altenburg, E., Hart, V., Gleason, K. & Carney, J. Consumer Reported Care Deferrals Due to the COVID-19 Pandemic, and the Role and Potential of Telemedicine: Cross-Sectional Analysis. JMIR Public. Health Surveill. 6, e21607 (2020).
Prachand, V. N. et al. Medically Necessary, Time-Sensitive Procedures: Scoring System to Ethically and Efficiently Manage Resource Scarcity and Provider Risk During the COVID-19 Pandemic. J. Am. Coll. Surg. 231, 281–288 (2020).
Bayindir, E. E., Messerle, R., Stargardt, T. & Schreyögg, J. Socioeconomic Disparities In Mortality And Health Care Use During The COVID-19 Pandemic: Evidence From Germany. Health Aff. 42, 1715–1725 (2023).
Bauer, A. & Weber, E. COVID-19: how much unemployment was caused by the shutdown in Germany? Appl. Econ. Lett. 28, 1053–1058 (2021).
Vonbun-Feldbauer, R. et al. Regionaler Pressemarkt Und Publizistische Vielfalt (Springer Fachmedien Wiesbaden, 2020). https://doi.org/10.1007/978-3-658-28965-2
Möhring, W. The need for local, everyday information. The importance of local journalism in the world of digital media. Journalism Res. 2, 54–60 (2019).
Bruns, A. Big Social Data Approaches in Internet Studies: The Case of Twitter. in Second International Handbook of Internet Research (eds Hunsinger, J., Allen, M. M. & Klastrup, L.) 1–17 (Springer Netherlands, Dordrecht, doi:https://doi.org/10.1007/978-94-024-1202-4_3-1. (2018).
Broersma, M. & Graham, T. Twitter as a News Source. Journalism Pract. 7, 446–464 (2013).
Bruns, A. & Moe, H. Structural Layers of Communication on Twitter. in Twitter and Society (eds. Weller, K., Bruns, A., Burgess, J., Mahrt, M. & Puschmann, C.) 15–28. Peter Lang US, (2014). https://doi.org/10.3726/978-1-4539-1170-9
Segev, E. Sharing Feelings and User Engagement on Twitter: It’s All About Me and You. Soc. Media Soc. 9, 20563051231183430 (2023).
Beisch, N. & Koch, W. 25 Jahre ARD/ZDF-Onlinestudie: Unterwegsnutzung steigt wieder und Streaming/ Mediatheken sind weiterhin Treiber des medialen Internets. Media Perspektiven. 11, 486–503 (2021).
Trezza, D. To scrape or not to scrape, this is dilemma. The post-API scenario and implications on digital research. Front. Sociol. 8, 1145038 (2023).
Mancosu, M. & Vegetti, F. What You Can Scrape and What Is Right to Scrape: A Proposal for a Tool to Collect Public Facebook Data. Soc. Media Soc. 6, 2056305120940703 (2020).
Gabrielatos, C. Keyness analysis: Nature, metrics and techniques. Corpus Approaches Discourse. 225–258. https://doi.org/10.4324/9781315179346-11 (2018).
Rayson, P. & Garside, R. Comparing corpora using frequency profiling. in Proceedings of the workshop on Comparing corpora – 1–6. Association for Computational Linguistics, Morristown, NJ, USA, (2000). https://doi.org/10.3115/1117729.1117730
Hardie, A. Log Ratio – an informal introduction. CASS (2014). https://cass.lancs.ac.uk/log-ratio-an-informal-introduction/
Sidarenka, U. P. T. S. et al. The Potsdam Twitter Sentiment Corpus. in Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC’16) (eds. Calzolari, N. 1133–1141 (European Language Resources Association (ELRA), Portorož, Slovenia, (2016).
Cieliebak, M., Deriu, J. M., Egger, D., Uzdilli, F. A. Twitter Corpus and Benchmark Resources for German Sentiment Analysis. in Proceedings of the Fifth International Workshop on Natural Language Processing for Social Media (eds. Ku, L.-W. & Li, C.-T.) 45–51. Association for Computational Linguistics, Valencia, Spain, (2017). https://doi.org/10.18653/v1/W17-1106
boyd, D., Golder, S., Lotan, G. Tweet Tweet, Retweet: Conversational Aspects of Retweeting on Twitter. in 43rd Hawaii International Conference on System Sciences 1–10 (IEEE, 2010). (2010). https://doi.org/10.1109/HICSS.2010.412
Salas, R. N. et al. Impact of extreme weather events on healthcare utilization and mortality in the United States. Nat. Med. 30, 1118–1126 (2024).
Rackow, B., König, H. H., Wall, M. & Konnopka, C. The interaction between air pollution, weather conditions, and health risks: a systematic review. Sci. Total Environ. 996, 180080 (2025).
Benoit, K. et al. quanteda: An R package for the quantitative analysis of textual data. J. Open. Source Softw. 3, 774 (2018).
Chollet, F. Keras. Preprint at https://keras.io (2015).
Abadi, M. et al. TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems. Preprint at (2015). https://www.tensorflow.org/
Xu, S. et al. Impact of the COVID-19 Pandemic on Health Care Utilization in a Large Integrated Health Care System: Retrospective Cohort Study. J. Med. Internet Res. 23, e26558 (2021).
Peine, A. et al. Telemedicine in Germany During the COVID-19 Pandemic: Multi-Professional National Survey. J. Med. Internet Res. 22, e19745 (2020).
Acknowledgements
We acknowledge assistance from several paid student assistants who performed annotations. ChatGPT4o was used while revising this manuscript.
Funding
Open Access funding enabled and organized by Projekt DEAL. Funded by Deutsche Forschungsgemeinschaft (DFG) - Grant number 458599862.
Author information
Authors and Affiliations
Department of English and American Studies, University of Hamburg, Hamburg, Germany
Julia Schilling & Robert Fuchs
Department of English, American and Celtic Studies, University of Bonn, Bonn, Germany
Julia Schilling & Robert Fuchs
Hamburg Center for Health Economics, University of Hamburg, Hamburg, Germany
Esra Eren Bayindir & Jonas Schreyögg
Authors
- Julia Schilling
- Esra Eren Bayindir
- Robert Fuchs
- Jonas Schreyögg
Contributions
J.S. (Julia Schilling) conducted the linguistic component of the study, including corpus construction, data preprocessing, sentiment and topic salience analysis, and figure preparation, and wrote the first draft of the manuscript. E.E.B. conducted the econometric modeling and regression analyses. R.F. and J.Sch. (Jonas Schreyögg) were responsible for project conception and study design, and provided critical feedback on the analysis and manuscript. All authors contributed to the interpretation of findings, revised the manuscript, and approved the final version.
Corresponding author
Correspondence to Julia Schilling.
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Schilling, J., Bayindir, E.E., Fuchs, R. et al. The impact of public discourse on healthcare utilization during the COVID-19 pandemic. Sci Rep 16, 31709 (2026). https://doi.org/10.1038/s41598-026-61014-7
Received: 12 August 2025
Accepted: 01 July 2026
Published: 10 October 2026
Version of record: 10 October 2026
DOI: https://doi.org/10.1038/s41598-026-61014-7
Keywords
来源:Nature 健康护理主题 · nature.com