Predictive model for long COVID in children 3 months after a SARS-CoV-2 PCR test

Children and young people (CYP) testing positive for SARS-CoV-2 are usually asymptomatic or have a low symptom burden at the time of infection compared to adults [1, 2]. Recent studies on post-COVID sequelae (also known as ‘long COVID’), however, have shown some adults and children can have persistent symptoms for months after acute infection [3, 4]. A recent systematic review of persistent symptoms following SARS-CoV-2 infection found most reported persistent symptoms were no more common in SARS-CoV-2-positive than in SARS-CoV-2-negative CYP, with only small increases in cognitive difficulties, headache, loss of smell, sore throat and sore eyes [5]. Similar to the successful use of predictive models for cardiovascular disease, e.g. in the United Kingdom (UK) [6] and the United States (US) [7], predictive models can help identify CYP at highest risk of experiencing persistent symptoms and direct them towards relevant care. This is particularly important during the pandemic when health services are under increased pressure [8]. A systematic review identified over 100 diagnostic and prognostic models for SARS-CoV-2, mainly relating to acute outcomes, e.g. mortality, intensive care unit (ICU) admission and length of hospital stay [9]. With the exception of two studies, however, most were considered low quality due to non-representative selection of controls, inadequate exclusions, high risk of model overfitting and unclear reporting [9]. Based on predictive model quality assessment tools [10] and model development guidelines [11], the two models mentioned above (the Jehi diagnostic model [12] and 4C mortality score [13]) and a third model (QCOVID [14]) are considered as higher quality predictive models for SARS-CoV-2 because of large sample sizes [15], appropriate modelling techniques [16] and suitable internal validation and reporting [11]. Of these three models, the 4C and QCOVID models were developed in adult populations (age≥18 years) whereas the Jehi model was developed in all patients who were tested for SARS-CoV-2 at all Cleveland Clinic locations in Ohio and Florida, US, regardless of age and included 11,672 patients (median age: 46.89 years among SARS-CoV-2 negatives; 54.23 years among SARS-CoV-2 positives).

There are very few predictive models for the potential long-term effects of SARS-CoV-2 infection, and those that exist have focused mostly on adults. Sudre and colleagues focused on identifying the characteristics and predictors of post-COVID sequelae in a sample of 4182 adults who reported testing positive for SARS-CoV-2 and found those experiencing more than five symptoms during the first week of illness were more likely to report ‘long COVID’ [17]. Recent large national cohort studies of CYP are consistent with the abovementioned systematic review [5], finding little difference in ‘long COVID’ symptom prevalence between SARS-CoV-2-positive and SARS-CoV-2 control CYP who either tested negative or did not have a test [4, 18]. As acute SARS-CoV-2 infection remains predominantly a mild infection in CYP and the cumulative incidence of infection increases, the incidence of post-COVID sequelae and the extent to which it is distinct from pandemic-related symptoms resulting from national lockdowns, school closures and social isolation is a critical factor in health policy decisions. We previously presented a model that predicted impairing symptoms in CYP [19], and here we aimed to update and internally validate the prediction model in CYP 3 months after a PCR test and to determine whether the impact of these predictors differed by SARS-CoV-2 infection status. The outcome examined here aligns with our previously described Delphi definition of long COVID [20].

We use data from the Children and young people with Long Covid (CLoCk) study: a national cohort study of SARS-CoV-2 PCR-positive CYP aged 11–17 years living in England who were matched at study invitation, on month of test, age, sex and geographical area, to SARS-CoV-2 test-negative CYP selected from the national testing database at Public Health England (now UK Health Security Agency (UKHSA)) [21]. Test-negative CYP who self-reported subsequently testing positive for SARS-CoV-2 were excluded [4].

Here we examine a previously described study subset that is broadly representative of the target population in terms of age, sex, geographical region and socio-economic status [4]. Briefly, from a total of 50,836 CYP who were approached, 7139 (3246 SARS-CoV-2 positive, 3893 SARS-CoV-2 negative) who completed the CLoCk questionnaire sent to them 3 months after their PCR test during January–March 2021 (median time between testing and questionnaire: 14.9 weeks [25th, 50th centiles: 13.1, 18.9]) were included. The questionnaire included demographic characteristics, elements of the International Severe Acute Respiratory and emerging Infection Consortium (ISARIC) Paediatric COVID-19 follow-up questionnaire [22] and the recent Mental Health of Children and Young people in England surveys [23]. CYP responded to 21 questions on physical symptoms at the time of testing (e.g. cough, tiredness, etc.). They rated their general physical and mental health before SARS-CoV-2 testing in two separate questions using a 5-category Likert scale. The prevalence of ‘very poor’ was low; therefore, for analysis, we recoded these variables into four categories (very poor/poor to very good). Quality of life/functioning before testing was measured via the EQ-5D-Y scale [24], and feelings of loneliness by the UCLA Loneliness scale [25]. The Index of Multiple Deprivation (IMD) was calculated from the CYP’s small local area level-based geographic hierarchy (lower super output area) at the time of the questionnaire and used as a proxy for socio-economic status. We examine IMD quintiles from most (quintile 1) to least (quintile 5) deprived (Table 1).

Of the 7139 CYP (3893 SARS-CoV-2 negative, 3246 SARS-CoV-2 positive) in our analytic sample, 26% (1860/7139) were of non-European origin, 62.9% (4493/7139) were female and there were more older than younger CYP (42.6% 16–17-year-olds vs. 31.0% 11–13-year-olds) (Table 1). Three months after their PCR test, 65.5% (2126/3246) of SARS-CoV-2 PCR-positives had at least one physical symptom (Table 2) and 25.2% (817/3246) had long COVID (i.e. at least one impairing symptom). This compares with 52.5% (2045/3893) and 18.5% (719/3893), respectively, in test-negative CYP.

To our knowledge, we have developed [19] and updated the first risk prediction model that uses self-reported information from CYP to estimate their probability of experiencing long COVID 3 months after SARS-CoV-2 testing. SARS-CoV-2 status, number of physical symptoms at testing, sex, age, ethnicity, self-rated physical and mental health, feelings of loneliness and four items from the EQ-5D-Y scale (all before testing) predicted long COVID 3 months later, with the impact of some predictors differing by SARS-CoV-2 status. We provide a risk calculator to predict CYP most likely to experience long COVID, to triage those who need support and for whom early intervention might be of greatest benefit. Importantly, our model has excellent predictive ability, calibration and discrimination. It enables us to answer important clinical questions such as ‘are those who have many symptoms during acute SARS-CoV-2 infection at greater risk of Long COVID than those without?’. The answer is ‘yes’ but our model provides a more nuanced answer by considering other factors.

Our goal was to provide a model that utilizes multiple factors (i.e. predictors) in combination, to accurately predict long COVID 3 months post-test. Importantly, our focus was not on whether included predictors are causal or not. Instead, the focus was the overall predictive performance of the model [33]. As such, we followed the guidelines to model building [27]. The large sample allowed flexible examination of the potential for relationships to differ by SARS-CoV-2 status and by the shape of the association without considerable concerns about overfitting. Model fitting statistics were extremely favourable and the use of a matched national cohort sample of test-positive and test-negative CYP is unique. Despite its internal validation, the model needs to be externally validated on other independent datasets and in different populations and settings prior to its wider application. Additionally, the model needs to be reassessed for experiencing long COVID beyond 3 months. It is possible many of the predictors stay the same but acknowledge there may be differences as the disease profile (and, therefore, predictors) changes over the course of the illness.

We acknowledge study limitations. The CLoCk study response rate (13.9%) is typical of surveys of this type [34] and is in line with other COVID-19-related studies [35, 36]. Importantly, the examined CYP are broadly representative of the target population in terms of important demographics such as age, sex and socio-economic status [4] as well as more generally of CYP aged 11–17 years living in England [37]. Baseline measures (at/or before testing) were subject to recall bias because they were not taken at the time of acute infection, and we were unable to assess whether symptoms waxed and waned between testing and the questionnaire. In addition, the possibility of selection bias in both directions (CYP more likely to participate if they have persistent symptoms, or less likely to participate if too unwell) among respondents cannot be ruled out. Furthermore, as the background epidemiological situation in relation to SARS-CoV-2 infection prevalence changes, there is a need to reassess possible differences in our model’s predictive value over time. Finally, caution is required for predictions based on data extrapolation/situations where there are only a very small number of observations for different predictor combinations.

To our knowledge, no other study has explicitly aimed to present a risk prediction model for long COVID [5, 38]. Moreover, the majority of previous studies lack a SARS-CoV-2 test-negative comparison group and so distinguishing long-term symptoms predicted by SARS-CoV-2 infection from background rates or pandemic-related effects remains a challenge [5]. More recent studies include control groups and, thus, broad comparisons can be made. Our finding that the odds of experiencing long COVID 3 months post-test was 1.48 times higher in SARS-CoV-2-positive compared to SARS-CoV-2-negative CYP is in line with findings from the LongCOVIDKidsDK study, where the SARS-CoV-2 test-positive group had 1.22 times higher odds of having at least one ‘Long COVID’ symptom lasting at least 2 months compared with the SARS-CoV-2 control group who either tested negative or never had a test [39]. We found both test-positive and test-negative CYP met the Delphi consensus definition of long COVID 3 months post-test with a difference of 6.7% between these groups. In contrast, in Borch et al., the prevalence of reported symptoms in CYP aged 6–17 years lasting more than 4 weeks was similar regardless of SARS-CoV-2 status (28% test-positives; 27.2% test-negatives/never had a test) [18]. Discrepancies in findings could be due to several reasons including differences in the symptom questions asked of the test-positive and test-negative/never been tested groups, timing of outcome (>4 weeks vs ~3 months), recruitment methodology, recruitment rates between test-positives and test-negatives and/or underlying prevalence levels in the countries at the time of the study. Our results are consistent with findings in adults, where the number of symptoms at onset [40] and female sex [41] were associated with ‘Long COVID’ and pre-existing diagnosis of depression/anxiety is over-represented in those with fatigue after SARS-CoV-2 infection [41].

Understanding which CYP are at risk of experiencing long COVID is important for individuals (e.g. in decision-making about whether to receive COVID-19 vaccination) and health service provision (e.g. for careful monitoring, early intervention and hopefully reduction in the burden of prolonged health problems). Using data from a large national matched cohort study, we updated our previously developed prediction model for experiencing long COVID 3 months after SARS-CoV-2 testing in CYP. Our model has excellent performance, and we hope it will serve as a useful tool for the early identification and management of CYP at risk of long COVID in the context of the current pandemic.