We estimated that before 31st July 2020 20.1% (19.2%, 20.7%) of identified COVID-19 cases in hospitals were likely to have been hospital-acquired infections and that within-hospital transmission likely contributed directly to 26,600 (mean, 95% range over 200 simulations: 25,900, 27,700) symptomatic infections, and a further 47,400 (45,000, 50,000) hospital-linked infections. These results are based on a 7 day cut-off for symptom onset from admission and prior to discharge for defining an identified hospital-acquired case.
Despite these levels of infection, we estimated hospital transmission to patients caused fewer than 1% of all infections in England in the first wave (prior to 31st July 2020). To some extent this reflects effective infection prevention within hospital settings with over 4 million non-COVID-19 patients being cared for in hospital settings during this period. However, the high proportion of hospital cases that were due to hospital-acquired infections is worrying as these are the most vulnerable members of our society and hence may have the most severe consequences. In addition, we did not account for the substantial proportion of asymptomatic infections in our analysis and thus, the impact of hospital transmission on the community epidemic is likely an underestimate [13].
This is the first study to estimate the total number of symptomatic hospital-acquired infections (not just the percentage of known cases that are hospital-acquired) and their wider contribution to community transmission prior to 31st July 2020. In particular, we found that the contribution of hospital-acquired infections to the epidemic likely varied over time, increasing in importance as community infections initially dropped, emphasising the need to determine where most infections are occurring at any one time during an epidemic. Analysis of subsequent waves of infection in England supports this wider contribution, finding that efforts to reduce in hospital transmission could substantially enhance the efficiency of potential community lockdown measures [23].
Our results show that relying on symptom onset > 7 days after admission in inpatients as a detection method for hospital-acquired SARS-CoV-2 will miss a substantial proportion (> 60%) of symptomatic hospital-acquired infections. This depends on the length of stay for non-COVID admissions but suggests that in many settings estimates of the number of infections due to transmissions in hospital settings will be substantial underestimates. For example, Read et al. [24] acknowledged that the estimated proportion of nosocomial infections during the first epidemic wave of COVID-19 in the UK that was based on symptom onset data, is likely to be higher if accounted for unidentified cases. This is particularly relevant for low-resource settings with short lengths of stay for non-COVID patients and which rely on symptom onset screening for SARS-CoV-2 infection.
An alternative cut-off, of say only 3 or 5 days from admission, would classify more infections correctly as hospital-acquired SARS-CoV-2 infections but would misclassify more community-acquired infections. Striking the right balance is difficult with a more reliable detection method being routine testing of patients, which will confirm symptomatic as well as detect pre-symptomatic and asymptomatic SARS-CoV-2 infections. However, even with screening on admission, symptomatic or not, and retesting 3 days after admission, a portion of infections will likely not be detected during inpatient stays due to short lengths of stay. Our estimates of the proportion of hospital cases that are due to hospital-acquired infection are higher than those from England wide studies [8, 24] and those from single hospital settings in the UK [3, 9, 25,26,27], as we estimate all hospital-acquired infections whether identified or not during their hospital stay. Our estimates of all infections are similar to previous modelling work using an SEIR model which estimates that nosocomial transmission was responsible for 20% (IQR 14.4, 27.1%) of infections in inpatients [28].
Our work implies that it may be effective to screen patients upon hospital discharge to detect infection, or to quarantine hospital patients on discharge to prevent ongoing community transmission: we estimate this would detect up to 40% of hospital-acquired infections that would become symptomatic (that would otherwise be “missed” in Fig. 3c). Hence, depending on the test sensitivity by time from infection, up to 70% of hospital-acquired infections could be detected. The onward community transmission from these infections may be especially important as community prevalence of SARS-CoV-2 infection decreases.
Currently, much more routine screening and testing is implemented in English hospitals contributing to the detection of infections prior to symptom onset or discharge [23, 29]. However, screening will need to be conducted with high frequency to avoid missing those infected prior to discharge, or to screen on, and for several days after, discharge. Our work is directly linked to the situation prior to August 2020 where little routine testing was in place and would be affected substantially by the new pandemic situation with new variants and vaccination. However, our conclusion that symptomatic screening of inpatients has limited efficacy in detecting nosocomial transmission is still highly relevant to support the need for ongoing regular screening of asymptomatic hospital patients and to emphasize potential missing infections.
Further work is needed to determine the precise risk of returning as a hospital case for those infected in hospitals. If our values (10–15%) are found to be conservative, then this percentage could increase substantially. If it were found to be higher, reflecting the poorer health of hospitalised patients and hence potentially increased susceptibility, then the proportion of hospital cases that are hospital-acquired could increase to 30–40%.
The interpretation of our results is limited by several simplifications. Firstly, we did not explicitly capture disease and hospital attendance variation by age. Future work could stratify our estimates to account for an older and more vulnerable hospital population. Secondly, we likely underestimated the total number of hospital-acquired infections as we modelled only those that progress to symptoms. While a non-negligible proportion of SARS-CoV-2 infections is likely to be asymptomatic [13], hospital-acquired infections were defined using the date of symptom onset in the UK. In addition, (a proportion of) symptomatic infections require medical care and therefore directly contribute to the hospital burden. We, thus, focussed on estimating the magnitude of under detection of these symptomatic hospital-acquired infections and their wider impact on community transmission.
Thirdly, we assumed a fixed number of four generations for onward transmission in the community to generate hospital-linked infections, and did not account for infections in healthcare workers, nor in the setting to which hospitalised patients were discharged to, such as long-term care facilities. The impact of onward transmission from hospital-acquired infections may be underestimated in this work since these settings may have high levels and large heterogeneity in onward transmission, or overestimated if four generations is longer than the average chain from recently hospitalised individuals. There is some data that, on average, this distribution is extremely right-skewed [30], but the likely different behaviour patterns of recently hospitalised individuals makes it hard to accurately predict length of transmission chains. Moreover, in our baseline scenario the number of secondary infections was usually less than one (time-varying R, Additional file 9) meaning that there would be diminishing numbers of secondary infections in each chain. Indeed, our sensitivity analysis shows that in the baseline, a further three generations contribute only a further ~ 50% of cases. However, with an increasing number of generations it becomes harder to contribute these linked cases to the transmission conditions in hospitals rather than to community transmission levels—our four generations of cases were chosen to be an indication of what may happen in the short time after hospital discharge.
Fourthly, we assumed that equal levels of infection control policies were in place in all NHS Trusts during this time period as we had no data to inform variation. Moreover, some of the “missed” cases may have been detected by community screening although there was little in place in England in this time (prior to August 2020). Finally, identification of hospital infection using CO-CIN relied on symptom onset date, which may be unreliably recorded potentially leading to bias in the patient population. While we cannot assess the biases, it is reasonable to expect that symptoms were recorded well in a clinical setting, and frequently (~ 65,000 patients included). An alternative definition of hospital-acquired infection reliant on the date of first positive swab would have its own limitations: patients could enter with symptoms and not test positive until more than a week into their stay [25].
We report our results around a baseline scenario and, despite including parameter sampling within multiple simulations, find a relatively small uncertainty range. Future work should build on our sensitivity analysis, which highlights the importance of understanding onward transmission (to accurately capture hospital-linked cases), to better understand disease and transmission heterogeneity and hence the importance of hospital settings to pathogen spread.