The COVID-19 pandemic and related mitigation measures have drastically changed our lives. Although political efforts have somewhat alleviated the economic and public health consequences of the pandemic, experts have warned that its long-term effects on mental health tend to be neglected [1-3]. Research conducted since the initial outbreak of the COVID-19 pandemic in China showed an increase in general stress [4] and a substantial increase in psychopathological symptoms that are frequently encountered in clinically relevant mood or anxiety disorders or both [5,6]. Preliminary evidence also suggests that survivors of COVID-19 appear to be at increased risk for mental health problems [7].

Worries and fear are centrally involved in COVID-19–related psychopathologies and problematic behaviors [8-11]. Schimmenti and coworkers [12,13] proposed a model to account for fear experiences during the COVID-19 pandemic. This model posits that several domains of fear (bodily, relational/interpersonal, cognitive, and behavioral) interact and contribute to the onset and perpetuation of COVID-19–related psychological distress through maladaptive, repetitive, and functionally impairing behaviors. One such behavior used to gain control over fear during the COVID-19 pandemic concerns compulsive searches for online health information, or “cyberchondria” [12,14,15].

Cyberchondria is defined as a poorly controlled pattern of searching for health-related information online, resulting in heightened health anxiety and other negative consequences (eg, interference with work or relationships and psychological distress), which can be functionally impairing and are associated with abnormal healthcare use [16,17]. Cyberchondria has been conceptualized as a multidimensional construct fueled by both anxiety and compulsivity-related factors [18] and described as a “transdiagnostic compulsive behavioral syndrome” [19], which is associated with health anxiety, problematic internet use, and obsessive-compulsive symptoms [20,21]. Cyberchondria is not included in the International Classification of Diseases 11th Revision or the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, and its defining features, etiological mechanisms, and assessment continue to be debated [22]. The upshot of this situation is that reliable data on the prevalence of cyberchondria in the general population are not available [19,23]. Nevertheless, preliminary data suggest that cyberchondria might be commonly encountered [24] and that it might be more frequent in patients with various medical conditions [25,26]. With regard to its psychological correlates, previous research has shown that cyberchondria is associated with low self-esteem, dysfunctional meta-cognitive beliefs, heightened anxiety sensitivity, and intolerance of uncertainty, as well as a tendency toward pain catastrophizing [19].

According to Starcevic et al [14], the COVID-19 context is likely to have contributed to the occurrence of cyberchondria or exacerbated it for several reasons: (1) there is a heightened perception of threat and the accompanying fear due to a recently identified and poorly understood disease; (2) uncertainty concerning the pandemic and the effectiveness of various mitigating measures (eg, lockdowns and vaccination) undermines attempts to cope with the situation; (3) the paucity of authoritative, trustworthy, and evidence-based health information further thwarts coping efforts; (4) the abundance of confusing, conflicting, unverified, and constantly updated information amplifies bewilderment; and (5) engaging in excessive online health information seeking cannot provide the necessary information and reassurance. These factors have been posited to increase fear and distress, thereby also increasing the perception of threat, further reducing effective coping with uncertainty and perpetuating online health searches. It is worth noting that the psychological model of cyberchondria during the COVID-19 described here [14] was developed at a time (March-May 2020) when the uncertainties surrounding the pandemic were at their maximum level and when the data for this research were collected.

In addition to this theoretical account, there is a growing number of empirical, mainly cross-sectional research reports focusing on various aspects of cyberchondria during the COVID-19 pandemic. Several important findings, in line with the psychological model proposed by Starcevic et al [14], have emerged from these studies. First, a strong relationship was found between cyberchondria and the fear of COVID-19 [27-30], with some studies reporting that cyberchondria predicts the fear of COVID-19 [29], other studies suggesting that the reverse might be true (ie, that the fear of COVID-19 predicts cyberchondria [30]), and yet other research reporting that both cyberchondria and health anxiety are risk factors for the fear of COVID-19 [27]. Second, several reports have confirmed the important role of intolerance of uncertainty during the pandemic, although the precise nature of its relationship with cyberchondria differs between studies [30-32]. Third, information overload was found to predict cyberchondria during the pandemic [33], whereas excessive and misleading information usually obtained through social media resulted in both cyberchondria and information overload [34]. Using a 2-wave longitudinal design during the initial outbreak of the pandemic in Europe, Jokic-Begic et al [35] showed that cyberchondria played a moderating role in the increase in the fear of COVID-19 between time 1 (when the first COVID-19 patients were diagnosed) and time 2 (when lockdown was introduced). Although these studies have improved our understanding of cyberchondria during the COVID-19 pandemic, much remains unknown about the psychological factors that contribute to the development of cyberchondria in the COVID-19 context.

In line with the assumption that cyberchondria is an important public health issue in the COVID-19 context [14,15], the objectives of this study were 2-fold. First, we investigated the levels of cyberchondria during the pandemic and compared them with the retrospectively assessed prepandemic levels of cyberchondria. Second, we aimed to identify the psychological factors that predicted cyberchondria during the pandemic. The selection of predictor variables was based on the psychological model of cyberchondria during COVID-19 [14], including the intolerance of uncertainty, COVID-19–related fears, health anxiety, and somatic symptoms. At the time the study was designed and conducted, the psychological model of Starcevic et al [14] was not yet published. Yet, some of the authors of this study were involved in its development and were thus able to capitalize on it for the selection of variables to be included in this study. In addition, we assessed impulsivity traits and attachment styles as predictor variables, because these psychological dimensions are potentially of relevance for behavioral patterns such as cyberchondria, which are characterized by diminished control and interpersonal difficulties [19]. To build a robust predictive model, this study used supervised machine learning–based regression models (elastic net regression).

Participants for this study were recruited using an online survey (created with Qualtrics), which was disseminated via social media (ie, Twitter, LinkedIn, Facebook, and Instagram). The study was also disseminated via the research networks of the authors and the scientific societies they are affiliated with. Data collection started on May 4, 2020, and ended on June 10, 2020, which corresponds to the first wave of the COVID-19 pandemic in Europe. At the time this study took place, French-speaking countries in Europe (France, Switzerland, Belgium, and Luxembourg) all implemented lockdown or semilockdown measures. The survey consisted of a questionnaire collecting demographic information (sex, age, education level, and country of residence) and information about socioeconomic circumstances during the first lockdown (eg, economic situation, housing, and employment status) and was followed by several instruments assessing various psychological and health-related constructs. The entire survey was administered in French. The survey software was set up in a way that participants could not skip any question, and therefore, we had no missing or incomplete responses in the final data set.

Some of the independent Italian data related to this project have been published elsewhere [10]. A list of all measures used in the online survey (including measures not considered here) is available from the Open Science Framework (OSF) [36]. All data, codes, and materials are available from the OSF link provided [36].

Inclusion criteria for the study were being at least 18 years of age and having a good understanding of French. No specific exclusion criteria were used. Sociodemographic characteristics of the participants are reported in Table 1. The sample consisted of 725 participants aged 18-77 (mean 33.29, SD 12.88) years, with females constituting the majority (416/725, 57.4%). Regarding a pandemic-related living situation, 5% (36/725) reported living with roommates during the lockdown, 20.4% (148/725) lived alone, 26.8% (194/725) lived with their children, 27.6% (200/725) lived with their parents, and 43.3% (314/725) lived as a couple. Most of the sample (626/725, 86.3%) assessed their housing situation as adequate during the lockdown. With regard to their financial situation, the majority of the sample (451/725, 62.2%) reported that they experienced no changes during the lockdown.

Participation was anonymous and voluntary. No compensation for completing the survey was provided. Participants were informed about the aims of the survey before they signed electronic informed consent. The study received approval from the Institutional Review Board for psychological research of the Kore University of Enna (UKE), in the framework of a joint Italian and Swiss research program on cyberchondria and COVID-19–related fears (code: UKE-IRBPSY-04.20.04).

Our first aim was to test whether the levels of cyberchondria increased during the pandemic in comparison with a retrospectively assessed cyberchondria, based on the CSS-12. As the CSS-12 scores in both response formats did not follow a normal distribution, we relied on nonparametric tests and computed Wilcoxon signed-rank tests for dependent samples. We also reported on the effects of gender, age, and education on the CSS-12 scores during COVID-19. The effect of gender was tested using the Mann-Whitney U test (nonbinary participants were not considered in this analysis due to their low number), and the effects of age and education were tested using Kruskal-Wallis tests (see Table 2 for more details).

Our second aim was to determine the factors that predicted cyberchondria during the pandemic, based on the psychological model elaborated by Starcevic et al [14]. Our predictive models focused on the CSS-12 subscales, which were most impacted by the COVID-19 pandemic (ie, those whose scores differed significantly from before the pandemic). Potential predictors for each model computed were selected based on their correlations with the dependent variable (ie, the CSS-12 subscales most impacted by the pandemic). Because we planned to apply a regression model, we did use Spearman correlations to select our predictors. Indeed, correlations can be used to quantify the dependence between our potential predictors and our dependent variable. Thus, all candidate predictor variables whose correlations with the dependent variable were ≥0.30 (which corresponds to a moderate effect size [55,56]) were retained and included in our predictive models. A series of predictive regression models were then computed based on a supervised machine learning approach.

Supervised machine learning approaches are generally defined as “a set of methods that can automatically detect patterns in data, and then use the uncovered patterns to predict future data” [57]. Traditional multiple linear regression models are limited in the sense that they rely on the entire sample to fit a model and test their accuracy. These models are also susceptible to bias and may be “overoptimistic” in terms of the variance explained or generalization to other independent samples. In contrast, the basic principle of the supervised machine learning approach is to shuffle the data (using a “seed,” which is a value set as a reference point to generate the randomization of the data) and then split them into 2 independent subsamples: one subsample is used to fit the model (train set, 60%-80% of the data), while the other is used to test the model’s accuracy (test set, 20%-40% of the data). Compared to the traditional regression approach, this method is generally considered to be more reliable and to produce more robust findings as the accuracy of the computed predictive model is derived from a new and independent sample with unknown variance [58,59]. Yet, such an approach needs a large sample to produce reliable findings, and another data-splitting strategy has been proposed in the context of supervised machine learning if the sample size is limited. This strategy is called cross-validation and involves a series of runs whereby the entire data set is split into several folds, which are all used as train and test sets [60]. In each run, a unique fold is used to determine the accuracy of the model computed, while the other folds are used to fit the model. Finally, each fold is used as a test set in one run and as a part of the train set in the other runs. The cross-validated score is obtained by computing a mean accuracy score based on the runs launched. This method is often used within the train set to “tune” the hyperparameters (a value that can be specified by the researcher) of a machine learning model. The fold used to compute cross-validation accuracy is called the validation set. Tuning a model consists of finding the hyperparameters that produce the best possible score on the validation set. When the hyperparameters are identified, the model is then refitted on the entire train set and its accuracy is evaluated using the test set. Nevertheless, this method has been criticized for promoting “overfitting,” in the sense that the model and its hyperparameters are too specific to the train set, thus potentially limiting its reproducibility [59].

An alternative method called nested cross-validation is depicted in Figure 1. This method bypasses the limitations of the classical cross-validation approach [59]. In nested cross-validation, an “outer loop” cross-validation is applied to split the data set into several folds to compute the overall accuracy. In each run, an “inner loop” cross-validation is performed to tune and validate the model by means of the folds used to fit the model (train set) in the outer loop. When inner-loop cross-validation is performed, the model is refitted based on the best hyperparameters identified on the folds used as train sets, and its accuracy is obtained from the fold used as the test set. In this study, we used the nested cross-validation method with hyperparameter tuning, and we repeated the procedure 25 times to achieve the most robust results possible, following guidelines provided by Vabalas et al [59] and Krstajic et al [61]. To select our machine learning model, we followed the flowchart provided by Scikit-learn’s documentation and concluded that elastic net regression is suited to our aim, considering our sample size and the number of variables used (sample N<100,000, and few features are used). Thus, the linear regression model elastic net, which combines ridge and lasso penalties, was used for our analyses [62]. A seed value of 1 was set for replicable results. In the Results section, we report a mean R² for each model computed as we obtained 1 R² per run (4 × 25 runs were computed; see Figure 1). We then computed the adjusted R² based on the formula 1 – [(N – 1)/(N – p – 1)] × (1 – R²), where p is the number of independent variables used in the model [63]. Finally, we compared the adjusted R² of the models using an independent t test.

Traditional statistics (Mann-Whitney U test, Spearman rank correlations, Kruskal-Wallis test, Wilcoxon signed-rank test, and multiple linear regression) were computed using R version 4.0.3 (R Foundation for Statistical Computing), and machine learning analyses (elastic net regression) were computed using the Scikit-learn version 0.24 Python module [64]. As most study variables did not follow a normal distribution, preliminary analyses were conducted to support the use of a linear supervised machine learning–based elastic net regression. We thus computed 1 traditional multiple linear regression and 2 generalized linear models (negative binomial and quasi-Poisson regressions). These 3 models all presented a significant P value (<.001) and showed similar results. Additional preliminary analyses are available from the OSF [36]. Internal consistency (Cronbach α) for all questionnaires used in the study was computed using Spearman rank correlations.

As shown in Table 2, a series of Wilcoxon signed-rank tests showed significantly higher scores during the pandemic on 2 facets of the CSS-12 (compulsion and distress subscales) than before the pandemic. Table 2 also shows significantly lower scores on the reassurance subscale of the CSS-12 during the pandemic and no significant differences before and during the COVID-19 pandemic on the excessiveness subscale of the CSS-12 and the total CSS-12 score. Gender, age, and education effects on the CSS-12 scores during COVID-19 are reported in Table 3. There were no gender differences with regard to the CSS-12 subscales and total scores. Age and education had some effect on the CSS-12 subscales and total scores, as shown in Table 3.

The 3 facets of the CSS-12, which proved to be affected by the COVID-19 context (distress, compulsion, reassurance) were considered in relation to our second objective, which was to identify the best predictors of pandemic-related cyberchondria. To select the variables to be included in the computed supervised machine learning–based models, the correlations with the 3 retained CSS-12 subscales were considered (the entire correlation matrix is reported in Table 4). As no correlation reached the threshold of ρ≥0.30 [55,56] for the reassurance subscale, this facet was not considered in further analysis. In contrast, potential predictor variables were identified for the distress and compulsion subscales.

A first supervised machine learning–based elastic net regression was computed for the distress subscale of the CSS-12. The following predictors were considered in the analysis: COVID-19–related fears (MAC-RF; ρ=0.515, P<.001), health anxiety (SHAI; ρ=0.491, P<.001), and intolerance of uncertainty (IUS-12; ρ=0.315, P<.001). As displayed in Table 5, the elastic net regression computed a mean R² of 0.344 (SD 0.059), and we obtained an adjusted R² mean of 0.333 (SD 0.06, 95% CI 0.321-0.345). The 2 most important predictors of the cyberchondria-related distress facet during the pandemic were COVID-19–related fears and health anxiety.

A second supervised machine learning–based elastic net regression was computed for the compulsion subscale of the CSS-12. The following predictors were considered in the analysis: COVID-19–related fears (MAC-RF; ρ=0.348, P<.001) and health anxiety (SHAI; ρ=0.355, P<.001). Both predictors included in the model (COVID-19–related fears and health anxiety) contributed similarly to the cyberchondria-related compulsion facet during the pandemic. As shown in Table 5, the elastic net regression computed a mean R² of 0.152 (SD 0.046), and we obtained an adjusted R² mean of 0.143 (SD 0.047, 95% CI 0.133-0.152), which is significantly lower than the one obtained for the model predicting the distress facet during COVID-19 (t₁₉₈=24.954, P<.001, 95% CI 0.175-0.205). The distress model contained 3 predictors, whereas the compulsion model contained only 2 predictors, which at least partly explains the lower explained variance for compulsion. It is, however, worth noting that the reported adjusted R² considered the number of predictors entered in the model.

This study aimed to determine whether the levels of cyberchondria changed during the COVID-19 pandemic and to identify the psychological predictors of cyberchondria during the pandemic. The results suggest that the facets of cyberchondria were affected during the COVID-19 pandemic following distinguishable patterns: although the levels of cyberchondria-related distress and compulsion increased, the levels of reassurance decreased. Using a supervised machine learning approach, we found that COVID-19–related fears (as assessed by the MAC-RF) and health anxiety (as assessed by the SHAI) were strong predictors of cyberchondria-related distress and compulsion during the pandemic.

An increase in the scores on the distress and compulsion subscales of the CSS-12 during the pandemic indicates higher levels of distress and greater interference with functioning, resulting from repeated online health searches. Scores on the reassurance subscale of the CSS-12 decreased during the pandemic, which suggests that online health searches were less likely to be conducted for the purpose of looking for medical professionals’ advice. This is possibly a consequence of either a sharply decreased availability of nonvital medical services during the first wave of the pandemic or the avoidance of medical facilities due to the fear of contracting COVID-19. Taken together, this pattern of results suggests that in the COVID-19 context, excessive online health searches do not provide reassurance, which may make these searches more distressing and cause impairment. Along the same lines, it is possible to speculate that the inability to obtain reassurance or necessary information via online health searches is also likely to increase the perception of threat and the accompanying fear of COVID-19, which may drive further searches.

These findings are in agreement with the theoretical model of cyberchondria during the COVID-19 pandemic [14]. Furthermore, they are in accordance with a suggestion that the “fear of not knowing” is a critical cognitive dimension of fear during the pandemic, which might increase distress and anxiety-related behaviors [12,13].

The scores on the excessiveness subscale of the CSS-12 did not show significant changes during the COVID-19 pandemic, which indicates that the general proneness to performing repeated online health searches does not necessarily change in the pandemic context. Likewise, total CSS-12 scores did not change during the pandemic, suggesting that the use of total CSS-12 scores in research may not reflect relevant or meaningful alterations in the patterns of problematic online health searches. This has implications for future research as the CSS is the most frequently used scale to assess cyberchondria [19,38], and studies conducted in the pandemic context have relied mainly on total scores either of the CSS-12 [29-31] or of the original CSS [65,66]. Therefore, it is advisable for future research on cyberchondria to always use scores on the CSS subscales in addition to total CSS scores. Furthermore, our findings raise concerns about the construct of cyberchondria, as assessed by various versions of the CSS, and support the notion that the issue of how best to assess cyberchondria needs to be revisited [38].

In view of our findings about the total CSS scores and scores on the specific CSS subscales, we specifically examined the predictors of the distress and compulsion facets of the construct of cyberchondria during the COVID-19 pandemic. The finding that COVID-19–related fears and health anxiety emerged as the strongest predictors of the distress and compulsion subscales of the CSS-12 supports the theoretical model of cyberchondria during the COVID-19 pandemic [14], as this model stipulates that the fear of COVID-19 is a key factor that drives online health searches in the pandemic context. A specific fear of COVID-19 and a more general propensity to be concerned about health and disease, as reflected in the construct of health anxiety, are likely to interact so that they mutually amplify one another. Our finding also confirms a significant relationship between health anxiety and cyberchondria that has been reported by numerous studies [20,21,37,67,68]. Moreover, other research has found a significant relationship between COVID-19–related fears and cyberchondria [27-30].

Other variables that were investigated in this study (somatic symptoms, intolerance of uncertainty, impulsivity traits, and attachment styles) did not emerge as strong predictors of either the distress or the compulsion facet of cyberchondria during the COVID-19 pandemic. Interestingly, the intolerance of uncertainty was a strong predictor only of the distress subscale of the CSS-12 but less so than COVID-19–related fears and health anxiety. Both previous research [30-32,67,69] and the theoretical model of cyberchondria during the COVID-19 pandemic [14] postulate a role for the intolerance of uncertainty in cyberchondria, but this role needs to be further investigated and better understood, alongside the impact of the fear of COVID-19 and health anxiety. With regard to impulsivity traits, their correlations with all subscales of the CSS-12 were the lowest, supporting the view that cyberchondria is better conceptualized as a behavior characterized by compulsivity or reassurance seeking [19,38] rather than impulsivity.

Our study comes with some specific limitations. First, we could have included additional predictor variables in our analyses. For example, maladaptive metacognitive beliefs have been associated with cyberchondria, both outside the COVID-19 context [67] and during the COVID-19 pandemic [29]. Yet, we selected our candidate predictor variables largely on the basis of the theoretical model of cyberchondria during the COVID-19 pandemic [14]. Other limitations include (1) our reliance on self-report instruments that may be affected by response biases (eg, social desirability, poor self-reflection abilities, and recall bias); (2) the cross-sectional nature of the study, which prevented us from investigating any causal relationships; (3) the self-selected nature of our sample, implying that it may not necessarily be representative of the general population (eg, our sample was mostly composed of highly educated individuals; see Table 1); and (4) the retrospective assessment of the prepandemic levels of cyberchondria.

This study contributes to the literature on cyberchondria in general and cyberchondria in the context of the COVID-19 pandemic in several ways. First, the facets of cyberchondria that pertain to distress and interference with functioning as a result of problematic online health searches became more prominent during the COVID-19 pandemic and were strongly predicted by COVID-19–related fears and health anxiety, supporting the theoretical model of cyberchondria during the COVID-19 pandemic [14]. Second, this is the first study of cyberchondria to use a supervised machine learning approach. Third, we showed that both cyberchondria as a multidimensional construct and its assessment need to be reexamined.

This study also confirms that cyberchondria is a public health issue of particular relevance during health crises, such as pandemics [14,15]. In such a context, it is important to identify factors that foster cyberchondria, because targeting these factors will contribute to efforts to prevent cyberchondria and tailor interventions for individuals displaying problematic online health searches.