The COVID-19 pandemic is one of the greatest health challenges that societies around the globe have ever experienced. A range of instruments and ways to measure factors related to COVID-19 and the pandemic have been described [1-7]. COVID-19 presents a challenge for public health in general, while health care workers (HCWs) are at particular risk of acquiring COVID-19 [8]. Several studies using online forms have found they can be useful for tracking disease activity in different locations, including workplaces [9,10]. However, these technological platforms require timely, persistent, and ongoing engagement to generate valid and representative surveillance data [1]. In the context of collaboration and the collection of collective health information, digital epidemiology and participatory surveillance techniques have been demonstrated to be tools with great potential for helping to detect health threats [11-16]. Many strategies that involve daily reporting of symptoms through the voluntary participation of individuals have reported successful results [17,18]. Participatory surveillance by patients has been shown to have a complementary role in detecting syndromic clusters for several epidemiological challenges, such as COVID-19, seasonal influenza, or high-risk mass gatherings [17-22]. The implementation of novel techniques represents an additional opportunity for the rapid analysis of big data based on machine learning, thereby acting as a complement to traditional disease surveillance systems.

The objective of this work is to describe a web-based participatory surveillance strategy among HCWs in two Swiss hospitals during the first wave of the COVID-19 pandemic.

From March 23, 2020, to August 23, 2020, a total of 127,684 SMS text messages were sent, generating 90,414 valid reports among 1004 participants, achieving a weekly average of 4.5 (SD 1.9) reports per user. Female gender (n=755, 75.2%) was more prevalent than male (n=249, 24.8%) among participants, reflecting the general HCW population in these hospitals. The median age was 39 years, with a mean of 40.2 (SD 11.3) years. Figure 1 shows the temporal distribution of symptoms of respiratory infection over the study period, using LOESS regression. In total, 1.49% (n=15) of participants reported a positive PCR result during the study period. The first peak of the bimodal curves clearly parallels the reference curve of individuals in the hospital who tested positive, representing the first COVID-19 wave in the region. The second peak appears between July 2020 and August 2020, with a much lower signal in the reference curve of individuals who tested positive.

Regarding anomaly detection over time, Figure 2 shows whether a signal of symptoms was expected (based in the past trends) or if it represented a positive or negative anomaly, meaning a significant increase or decrease in the frequency of recorded symptoms. Table 2 indicates the change points that were statistically significant, including the difference observed when compared with the expected amount. The positive anomalies happened in three different periods; two of them occurred during the highest activity of the first wave and the third occurred between July and August, representing a possible second wave. However, as mentioned above, no second (or third) wave was seen in the reference curve.

A correlation matrix between symptoms and a positive PCR test result for SARS-CoV-2 is shown in Figure 3, while in Figure 4, the significance matrix shows the positive and negative correlations, as well as the nonsignificant ones. The symptom with the strongest correlation with a positive PCR result was loss of taste. Conversely, symptoms such as red eyes or runny nose were negatively associated with a positive test (Table 3).

Finally, Table 4 shows the summary results from a random forest algorithm that was used to classify participants into SARS-CoV-2 positive and negative cases based on their indicated symptoms. The area under the ROC curve shows reasonable performance of the classification tree, with an accuracy of 88% for the training data and 89% for the test data (Figure 5). Nevertheless, while the prediction matrix showed good specificity (80.0%), sensitivity was low (10.6%; Table 5). Figure 6 shows the importance of symptoms and their capacity to predict the expected outcome based on the random forest algorithm, considering a P value of <.05. Loss of taste and limb/muscle pain were the most important variables for prediction of a positive result, while runny nose and red eyes were negatively correlated with the same outcome. Fever was a very weak predictor of a positive result.

This study demonstrates the use of digital surveillance to monitor COVID-19 activity among HCWs. Loss of taste was the symptom that was most aligned with COVID-19 activity at the population level. At the individual level, using machine learning–based random forest classification, reporting loss of taste and limb/muscle pain as well as absence of runny nose and red eyes were the best predictors of COVID-19. The main strengths of the study are its high response rate and the comparison to a reference curve, which was based on documented PCR results in the same population.

Syndromic surveillance through participatory surveillance has been shown to be a feasible strategy to monitor COVID-19 activity [33], and is considered an important measure to inform the public health response to this pandemic [34]. Considering that engagement is a key element of a successful platform, our study—with an average response of 4.5 answers per week—has an excellent basis to produce valid and representative results. This high rate of engagement and participation is extraordinary when compared to other platforms [13,17,18,33], especially over a period of 5 months [35]. The easy-to-use survey, the defined population of HCWs from two different hospitals, and the regular interaction with study participants are potential reasons for this high response rate. It remains to be seen if these engagement indexes can be maintained when the study is scaled up to larger communities.

The temporal distribution of symptoms followed the trends represented in the first wave of COVID-19 in Switzerland [36,37]. However, the signals detected in July were not due to COVID-19, as shown by the reference curve. Interestingly, several HCWs tested positive for rhinovirus during this time period, suggesting that this was the reason for this wave. Of note, loss of taste, the most specific symptom of COVID-19, did not increase during this second wave.

Several other studies have shown that loss of taste is a good proxy for COVID-19 [38-41]. Although the specificity of this symptom is excellent, only about 20% of patients report loss of taste [42]. We conclude that the detection of loss of taste is very helpful to interpret findings at the population level, but less so at the individual patient level because of its low prevalence. The second most important positively associated symptom in our analysis was limb/muscle pain, which has also been noted by others [43]. Remarkably, runny nose and red eyes were very important negative predictors of COVID-19; this finding is particularly useful for when surveillance is performed during allergy season. However, both the sensitivity and specificity of a symptom depend on the background activity of other infections and allergies and might therefore be subject to change. The validity of a symptom may also change due to genetic adaptations in the dominant SARS-CoV-2 strain. During the study period, none of the new variants of SARS-CoV-2 (eg, B.1.1.7/Alpha) were circulating in Switzerland. Therefore, the symptoms described here cannot necessarily be extrapolated to a different circulating SARS-CoV-2 variant. However, syndromic surveillance through participatory surveillance may allow for the detection or validation of a different clinical presentation emerging from a new circulating strain. Indeed, a recent study describes small differences in COVID-19 symptoms in the general population in the United Kingdom depending on the variant [44].

Our study has a number of limitations. First, it was performed outside influenza season. Because influenza more often presents with constitutional symptoms than other respiratory viruses, distinguishing influenza from COVID-19 by analysis of symptoms is difficult. Second, we relied on participants self-reporting their symptoms, a method that is prone to bias. Third, generalizability of our data is limited because only one-fifth of the HCWs from our hospitals participated in the study; in addition, the spatial component could not be explored due to these same reasons. At the same time, this would be a very important parameter for evaluating whether SARS-CoV-2 is being regionally distributed, which would be useful to form a complete picture for disease surveillance purposes. The application of classification techniques based on machine learning, such as random forest classification, has its own limitations, as a large number of trees can make the algorithm too slow and ineffective for real-time predictions. In general, these algorithms are fast to train, but quite slow to create predictions once they are trained. A more accurate prediction requires more trees, which results in a slower model.

Nevertheless, we deem the presented surveillance tool highly useful in monitoring and predicting COVID-19 activity among our HCWs. Currently, we have expanded our HCW cohort to include over 5000 participants from over 20 institutions [45]. The analysis of data from different institutions will allow us to detect the clustering of cases in certain institutions, which might trigger targeted intervention measures in affected health care institutions. Additionally, these data allow for the detection of symptomatic HCWs who were either not tested or had a false-negative PCR result, and also for the discrimination of symptoms caused by SARS-CoV-2 from symptoms caused by other viruses, such as influenza. Further questions, which we aim to answer with the surveillance data generated in this larger cohort, include how long HCWs with documented SARS-CoV-2 infection (or vaccination) are protected against reinfection or how the emergence of viral variants might change the symptomatology of COVID-19.