The predictive role of symptoms in COVID-19 diagnostic models: A longitudinal insight

Abstract

To investigate the symptoms of SARS-CoV-2 infection, their dynamics and their discriminatory power for the disease using longitudinally, prospectively collected information reported at the time of their occurrence. We have analysed data from a large phase 3 clinical UK COVID-19 vaccine trial. The alpha variant was the predominant strain. Participants were assessed for SARS-CoV-2 infection via nasal/throat PCR at recruitment, vaccination appointments, and when symptomatic. Statistical techniques were implemented to infer estimates representative of the UK population, accounting for multiple symptomatic episodes associated with one individual. An optimal diagnostic model for SARS-CoV-2 infection was derived. The 4-month prevalence of SARS-CoV-2 was 2.1%; increasing to 19.4% (16.0%-22.7%) in participants reporting loss of appetite and 31.9% (27.1%-36.8%) in those with anosmia/ageusia. The model identified anosmia and/or ageusia, fever, congestion, and cough to be significantly associated with SARS-CoV-2 infection. Symptoms' dynamics were vastly different in the two groups; after a slow start peaking later and lasting longer in PCR+ participants, whilst exhibiting a consistent decline in PCR- participants, with, on average, fewer than 3 days of symptoms reported. Anosmia/ageusia peaked late in confirmed SARS-CoV-2 infection (day 12), indicating a low discrimination power for early disease diagnosis.

Keywords: coronavirus; longitudinal data; symptoms dynamics.

Figure 1.

Age distribution in the study sample compared to that of the UK population, stratified by gender and ethnicity.

Figure 2.

Proportions of participants with specific symptoms, overall, and stratified by PCR status, as shown in the Supplementary Material. For example, overall, 16.9% of all participants reported runny nose at least once but the figure is much higher (72.6%) among PCR+ contrasting with 15.7% among PCR−.

Figure 3.

Predicted probabilities of PCR+ status, stratified by the presence of specific symptoms, and their 95%CIs. Predictions related to each specific symptom are unadjusted for the others and are based on a binary regression with robust standard errors accounting for multiple episodes with events associated with a participant. For example, in participants with loss of taste or smell, regardless of the presence or absence of other symptoms, the probability of a positive PCR test is 0.319 (31.9%).

Figure 4.

Predicted mean of number of days specific symptoms were reported during an episode and their 95%CIs. The red values (PCR+) are referred to the left axis and the blue values (PCR−) are referred to the right axis. The analysis is restricted to symptomatic participants only. For example, for those participants reporting cough as part of an episode, the mean of the number of days was 6–7 days in PCR+ participants and 2–3 days in PCR−.

Figure 5.

Daily probabilities of reporting specific symptoms starting with the first report conditioned on PCR+ participants and their corresponding illness episode, that is, ignoring the symptomatic episodes associated with these participants which were PCR-. Non-parametric methodology was used to capture the shape of the individual longitudinal daily reports.

Figure 6.

Daily probabilities of reporting specific symptoms starting with the first report using PCR- symptomatic episodes across all participants.

Figure 7.

Probabilities of daily occurrences of various symptoms have similar magnitude in both PCR+ and PCR− groups on the first reporting day whilst they peak up later during illness evolution in PCR+ patients and decline in those PCR−, also reflected in previous figures 5 and 6.

Figure 8.

Effect (OR) of reporting a specific symptom for 3 days during an episode, irrespective of other symptoms reported during that episode.

Figure 9.

Discrimination power of individual symptoms based on the temporally ordered reports restricted to the first 1, 2, 3 to longer than 15 days after the symptomatic illness episode starts.

Figure 10.

Estimated discrimination power of each classifier. The plot and the AUC estimate follow a maximum likelihood ROC-weighted regression analysis uncontrolled for age and ethnicity.

Figure 11.

Effect of age and ethnicity on the ROC curve and subsequently on discrimination power associated with each classifier in the model. The colours indicating specific symptom are similar to those displayed in Figure 10.