Baseline demographics
Figure 1 shows the patient's selection process. In total, 120 patients were included in the study. The median age of patients was 56 years (IQR 43 to 68). The median length of hospitalization was 10 days (IQR 9 to 29). Table 1 shows the distribution of the baseline characteristics grouped into ward and ICU patients. According to the CCI, 40% of patients belonged to the moderate or severe category. The most common comorbidities were diabetes mellitus and hypertension. There was a high incidence of new-onset diabetes, affecting 5% of patients. Compared to patients in the ward, the patients in ICU were significantly older and more likely to be male.
Table 1 Baseline characteristics of COVID-19 patients during the hospital stay stratified by ICU admission.
Symptom profile
Table 2 shows the prevalence of the PCC at different time points. The symptom prevalence decreased during the one-year follow-up period across the spectrum. Throughout the one-year follow-up period, the most common PCCs were constitutional, followed by respiratory and neuropsychiatric symptoms. At the end of the year, the prevalence of PCC in the cohort was 32.8%, with the prevalence of constitutional symptoms at 16%, respiratory symptoms at 8.4%, and neuropsychiatric and musculoskeletal symptoms at 2.5% each. Among constitutional symptoms, fatigue, loss of appetite, and weight loss were common. It should be noted that the prevalence of fatigue sharply decreased over time, from 67.5% at two weeks to 7.6% after one year. On the other hand, loss of appetite and weight loss didn’t show such a sharp decline and persisted for up to one year. With respect to the respiratory symptoms, both dyspnea and cough were common, and while they decreased over time, 5.9% of patients still experienced dyspnea at the end of one year. With respect to neuropsychiatric symptoms, sleep disturbance was the most common, and while it decreased somewhat over time, 5% of patients still experienced it after one year.
Baseline predictors for PCC
Figure 2 shows the risk of developing PCC at different time points depending on baseline characteristics. The figure also shows the risks of being readmitted to the hospital after discharge due to PCC. Not only do PCCs become less frequent over time (Table 2), but predicting them from baseline characteristics becomes more difficult the further away from the baseline the prediction is made. In fact, none of the demographic variables are a significant predictor of PCC after two weeks. While drug use during the hospital stay or given at discharge is a strong and significant predictor, the simple explanation is that patients with severe COVID-19 conditions are more likely to be given drugs and are also more likely to develop PCC. For this reason, drug use might be a good predictor but cannot be considered a risk factor for PCC. COVID vaccinations reduced the risk of persistence at one year and the risk of rehospitalisation. However, since 95% of the participants were vaccinated with at least two doses, the confidence interval for the difference between vaccinated and unvaccinated participants is rather wide (OR 0.2, 95% CI 0.03--4.8).
Biomarker predictors for PCC
Figure 3 shows the diagram of the Bayesian network. The estimated accuracy from cross-validation is 91%. All patients who were hospitalized had constitutional symptoms. Among patients with constitutional symptoms, the odds ratio of being re-hospitalized was 2.34 for respiratory symptoms, 1.77 for neuropsychiatric symptoms, and 2.95 for other symptoms (cardiovascular, gastrointestinal, infections, and dermatological). The relationships between symptom groups and the risk of rehospitalisation were all significant according to Fisher’s exact tests (p < 0.01).
Table 2 Frequency of post covid conditions in symptom groups at two weeks, six weeks , and one year after hospitalization
Location: Below the sub heading ‘symptom profile’ in results section.
In the Bayesian network model, all symptom groups and their interactions with the constitutional symptoms were highly significant (p < 0.01) predictors for the risk of rehospitalisation. The relationships between the biomarkers and the symptom groups are summarized in Table 3 and Table 4. Table 3 lists the biomarkers in the network that have a direct effect on the risk of the development of symptom groups in the Bayesian network. Similarly, Table 4 lists the modified effects of the biomarkers on the symptom groups in the network. CRP is the most common effect modifier, modifying the effects of eosinophils, urea, and potassium on neuropsychiatric symptoms and the effects of magnesium and Lactate Dehydrogenase (LDH) on other symptoms. The two other effect modifiers in the network are the CCI and sex: with CCI modifying the effect of LDH on neuropsychiatric symptoms and the effect of serum albumin and D-Dimer on other symptoms, and sex modifying the effect of serum globulin on constitutional symptoms.
Table 3 Estimated effects (mean differences) in the lab measurements depending on the presence/absence of a particular symptom group. Additionally, the mean and the standard deviation of each lab measurement in the study sample is shown.
Table 4 Estimated effects (mean differences) of lab measurements on the presence/absence of a particular symptom group in each category of an effect modifier. The interaction effect measures the difference between the low-risk and high-risk of the effect modifier. Additionally, the mean and the standard deviation of each lab measurement in the study sample are shown.
Figure 4 illustrates the effect of serum globulin and how this effect is modified by sex. Figures 4a and 4b show that neither serum globulin nor sex can predict the outcome independently. Figure 4c shows that serum globulin is higher in females than in men on average. However, stratifying by sex in Figure 4d, we see that serum globulin is significantly different between the 4 groups and that for men, higher globulin levels are associated with constitutional symptoms. This influence of serum globulin is illustrated in Figure 4e, where globulin levels are now discretized into two groups, less than 2 g/dL, and more than 2 g/dL, and the two globulin groups are predictive for men.