Health systems, medical providers, bioinformaticians, and researchers worldwide are working tirelessly to understand, contain, and ameliorate the COVID-19 pandemic. During this health emergency, clinicians have anecdotally noted an unusual number of patients with respiratory complaints at the end of 2019 and early 2020, well before COVID-19 was officially categorized by the World Health Organization (WHO) as a pandemic [1]. It is unclear whether such anecdotal reports are correct or the result of hindsight bias. If correct, the excess could represent typical variation in disease patterns. Alternatively, the excess, especially if it is significantly above prediction intervals based on historical data, could represent undetected and early COVID-19 cases prior to established clinical awareness and testing capabilities for the virus.

In the past decade, there has been widespread adoption of electronic health records (EHRs) in the United States. However, there have been limited efforts to date to leverage EHRs to support the delivery of high-value medical care or otherwise improve the delivery of health care services [2]. Using EHR data to model and forecast trends has the potential to improve resource management and the preparedness of health systems [3-7], which in turn could improve the quality of medical care. In particular, EHR data paired with analytical tools can potentially identify unusual trends in health care delivery that can alert clinicians and public health experts to critical changes in disease patterns.

The purpose of this paper is to use discrete raw EHR data to evaluate whether there was an excess of patients presenting with symptoms and diseases suggestive of COVID-19 in the months prior to the first known COVID-19 cases in the US health system in March 2020, using words found in chief complaint fields and International Classification of Diseases (ICD) codes from hospital discharge diagnoses. We analyzed 5 years of data from a large Los Angeles–area health system using time-series methods to address whether there was an excess number of patients presenting for complaints of cough, or hospitalizations for respiratory ailments. These methods highlight how health care analytics coupled with EHR data can be harnessed for disease surveillance. In particular, surveillance starting from the larger outpatient setting, which is often the tip of the iceberg, can provide an early warning of a public health emergency before patients fill hospital intensive care units and deaths accumulate.

The study included data from July 1, 2014, to February 29, 2020, from UCLA Health, a large health system with over 2.5 million total outpatient visits and 3 hospitals (UCLA Medical Center Santa Monica, Resnick Neuropsychiatric Hospital at UCLA, and Ronald Reagan UCLA Medical Center). Health system utilization data during the winter season, from December 1, 2019, to February 29, 2020, the months prior to increased public awareness of COVID-19 in the United States, were evaluated using the previous 5 years as the comparison period in a time-series analysis. Data were collected using SQL reports from Epic Clarity production databases supporting the EHRs used throughout UCLA Health.

Analyses included three different care settings: outpatient clinics, emergency departments, and hospital. All primary, specialty, and urgent care outpatient visits were considered and searched for the word “cough” within the reason for visit, further examining the percentages of patients presenting with cough in the current winter season with forecast predictions based on data for the preceding 5 years. All patient visits to emergency departments for cough and data on patients hospitalized with acute respiratory failure for the current winter season were separately examined with the corresponding data for recent years in the same manner (see Table 7 in Multimedia Appendix 1 for a list of ICD codes).

Seasonal autoregressive integrated moving average (SARIMA) models were applied on weekly data in SAS (SAS Institute) from July 1, 2014, through November 30, 2019, to forecast data for December 2019 through February 2020 (Tables 1-6 and Figures 1-3 in Multimedia Appendix 1). These models take into account seasonal effects [8,9] and use maximum likelihood to estimate model parameters based on historical data from July 1, 2014, to November 30, 2019. A winter season is defined as the time period from December 1 to the last day of February the subsequent year. Using the SARIMA model, a forecast of the 2019/2020 winter season was provided. Specifically, a SARIMA(1,0,1)x(1,0,1)₅₂ model was used to analyze the outpatient time series, a SARIMA(1,0,3)x(1,0,1)₅₂ model was used to analyze the emergency department data, and a SARIMA(1,0,1)x(1,0,1)₅₂ model was used to analyze the inpatient data for patients with acute respiratory failure. The autoregressive and moving-average orders were identified by both SCAN and extended sample autocorrelation function (ESACF) methods [10].

The 95% prediction intervals for the forecast allowed an assessment of whether the observed data at weekly intervals for the 2019/2020 winter season were outside of the time-series prediction interval.

For example, for the outpatient clinic visit data, the following SARIMA(1,0,1)x(1,0,1)₅₂ model was used (model parameter estimates are shown in Table 1 in Multimedia Appendix 1):

Where X_t is the percentage of visits in week for which cough was a recorded symptom; B is the backshift operator, BX_t = X_t-1; is a white noise process of uncorrelated random variables with mean 0 and variance ; and Z_t is an indicator variable for the months of December to February.

Excess cases per 1000 visits for the 2019/2020 winter season compared to previous seasons were estimated using three methods (Tables 8-11 in Multimedia Appendix 1). First, for a conservative estimate of excess cases, the maximum of the five previous seasons was subtracted from the current season. Second, the average of the five previous seasons was subtracted from the current season. For both methods, the excess percentage was multiplied by the total number of patient visits in the current season to estimate the excess cases for each month. For the third method, the upper limit of the time-series 95% prediction interval was subtracted from the observed rate for each of the 13 weeks in the 2019/2020 season to estimate the weekly excess percentages. The weekly excess percentage was multiplied by the weekly patient visits and aggregated to estimate the excess cases.

To visualize the data, the local regression (LOESS) technique was used to smooth daily data with a smoothing span of 20% (see Figure 4 in Multimedia Appendix 1 for scatter plots of daily data) [11], with analyses performed in R (R Foundation for Statistical Computing) [12] and figures generated in Microsoft Excel (Microsoft Corp).

Two sensitivity analyses were performed. First, patient insurance status was considered to see if variation in insurance could explain trends over time in the outpatient and emergency department data; the analysis was repeated using only those outpatient clinics that existed for all years of the study period to ensure that changing case mix did not confound our results. Second, patients hospitalized with a broader set of respiratory illnesses were investigated: patients hospitalized with ICD codes used in a study of respiratory tract illnesses associated with influenza [13], and those with any pneumonia (see Table 7 in Multimedia Appendix 1 for a list of ICD codes). A SARIMA(2,0,2)x(1,0,1)₅₂ model was used to analyze the inpatient data for “any respiratory tract disease,” and a SARIMA(2,0,1)x(1,0,1)₅₂ model was used to analyze the inpatient data for “pneumonia.” Institutional Review Board approval was obtained (UCLA number 20-000528).

A significantly higher number of patients presented to outpatient clinics and emergency departments in this health system with a complaint of cough starting the last week of December 2019 and continuing through January and February 2020. These findings translate into hundreds of additional patients seeking outpatient medical attention in this health system for the respiratory symptom of cough during the 2019/2020 winter season. A significant excess in the number of patients hospitalized with acute respiratory failure during this same time period was also noted. It is possible that some of this excess represents early COVID-19 disease before clinical recognition and testing, information that may help epidemiologists better understand the spread of this pandemic. If only some of these excess visits are due to COVID-19, this could still represent community spread of SARS-CoV-2 during that time because a substantial proportion of individuals infected have no symptoms or mild symptoms and do not seek medical care [14,15], making cryptic spread of the disease within a community likely.

EHRs are widely adopted but have not been used to their fullest potential to deliver high-value care [16-18]. This work demonstrates the potential of using EHR data for symptom or disease surveillance. A strength of this study is the use of raw EHR data that are already collected in most health care systems to determine whether patients were presenting with reports of “cough” at excess numbers in the months before the first known case of COVID-19 in this health system. In addition, this study considers data from three separate locations of patient care contained in the EHR: the outpatient setting, the emergency department, and the hospital. For many diseases, data from the outpatient setting can provide an early warning to emergency departments and hospital intensive care units of what is to come. By leveraging time-series analysis to calculate prediction intervals of expected patients, outlier numbers of patient visits can be quickly identified.

While asymptomatic transmission and community spread of COVID-19 are possible explanations for the observed excess patient encounters, other reasons and limitations need to be considered. The study was performed in a single health system and we only searched for the word “cough.” Although the term “cough” is possibly more specific to COVID-19 than other symptoms such as “fever” or “aches,” this search method has imperfect specificity and sensitivity as it does not include the full spectrum of COVID-19 symptoms [19-25]. The health system patient mix could have changed over the study period, but sensitivity analyses did not find evidence that this affected the findings. It is possible that the findings are due to lung injury from e-cigarettes (vaping), but this explanation is doubtful because the Centers for Disease Control and Prevention reported a continued decline after September 2019 [26].

Another limitation is not knowing for certain whether and what percentage of excess patient visits were due to influenza. An increase in influenza-positive test results and emergency department visits for influenza-like illnesses was noted in Los Angeles County and the United States during the 2019/2020 winter season when compared with prior years [27,28]. The incidence of influenza-like illness symptoms in the United States peaked earlier in 2019 when compared with previous years [28], with the 2019/2020 season peak at a similar level as the 2017/2018 season peak. This suggests that some of the observed incidence was due to influenza rather than COVID-19. However, the analysis shows an excess of patients above the forecast 95% prediction interval based on five previous seasons.

It is plausible that some of the excess visits might be due to SARS-CoV-2 as studies of rapid sentinel surveillance and genome sequencing suggest community transmission of SARS-CoV-2 much earlier than initially thought. Studies in the United States and France found evidence for cryptic spread of the virus as early as December 2019 or January 2020, before community surveillance was actively implemented [29-32]. Especially early in outbreaks, existing methods for case identification may not capture incident infections; thus, novel and complementary methods using EHR data such as those reported here may play an important role.

Heightened media attention regarding the coronavirus pandemic could influence patients to seek medical care for cough-related concerns and should be considered [33]. Such data can complement these EHR data approaches. Using data from Google Trends [34], the popularity of “cough” as a search term in the United States was slightly higher from late December 2019 through January 2020 when compared with the average national popularity in the previous 4 years (Figure 8 in Multimedia Appendix 1). Popularity of searches for “cough” increased substantially in mid-March, corresponding to the sharp increase in “coronavirus” Google news searches. Although there was almost no mention of a COVID-19–type illness by United States media in December 2019 and little mention in January 2020, substantial media attention was present in February 2020. Therefore, this limitation is pronounced for February 2020, but less concerning for the earlier months.

In summary, health system analytics combined with EHR data can be harnessed to quickly identify changes in underlying patient populations. This study identified a significant excess of patients with COVID-19–like presentations starting the last week of December 2019 and continuing through February 2020, a time period before the availability of testing or providers considered clinical diagnoses of COVID-19. A unique feature of this study is the evaluation of three different stages of health care settings, which expands surveillance beyond just reporting the number of patients in emergency departments or hospital intensive care unit beds to include consideration of 9.5 million outpatient clinic records. The Centers for Disease Control and Prevention Outpatient Influenza-like Illness Surveillance Network (ILINet) currently monitors patients presenting with fever and a cough and/or a sore throat in the United States [35]. Data from the outpatient clinic setting is usually a harbinger of what is to come for hospital emergency departments and intensive care units.

Harnessing larger electronic health data systems to monitor outpatient visits for the growing and diverse set of symptoms associated with COVID-19 should be considered [21-25].

This SARS-CoV-2 pandemic highlights the urgent need to support the development of agile health care analytics that enable real-time symptom and disease surveillance [36]. Lessons learned from this pandemic will hopefully lead to better preparation and the ability to quickly provide warnings and track the next pandemic.