key: cord-0699290-s0zwmehf
authors: Stirrup, O. T.; Hughes, J.; Parker, M.; Partridge, D.; Shepherd, J. G.; Blackstone, J.; Coll, F.; Keeley, A. J.; Lindsey, B. B.; Marek, A.; Peters, C.; Singer, J. B.; The COVID-19 Genomics UK consortium,; Tamuri, A.; de Silva, T. I.; Thomson, E. C.; Breuer, J. C.
title: Rapid feedback on hospital onset SARS-CoV-2 infections combining epidemiological and sequencing data
date: 2020-11-15
journal: nan
DOI: 10.1101/2020.11.12.20230326
sha: 868a01b74f1a874062c20a61fc83bc0d44a1dcff
doc_id: 699290
cord_uid: s0zwmehf

Background: Rapid identification and investigation of healthcare-associated infections (HCAIs) is important for suppression of SARS-CoV-2, but the infection source for hospital onset COVID-19 infections (HOCIs) cannot always be readily identified based only on epidemiological data. Viral sequencing data provides additional information regarding potential transmission clusters, but the low mutation rate of SARS-CoV-2 can make interpretation using standard phylogenetic methods difficult. Methods: We developed a novel statistical method and sequence reporting tool (SRT) that combines epidemiological and sequence data in order to provide a rapid assessment of the probability of HCAI among HOCI cases (defined as first positive test >48 hours following admission) and to identify infections that could plausibly constitute outbreak events. The method is designed for prospective use, but was validated using retrospective datasets from hospitals in Glasgow and Sheffield collected February-May 2020. Results: We analysed data from 326 HOCIs. Among HOCIs with time-from-admission [≥]8 days the SRT algorithm identified close sequence matches from the same ward for 160/244 (65.6%) and in the remainder 68/84 (81.0%) had at least one similar sequence elsewhere in the hospital, resulting in high estimated probabilities of within-ward and within-hospital transmission. For HOCIs with time-from-admission 3-7 days, the SRT probability of healthcare acquisition was >0.5 in 33/82 (40.2%). Conclusions: The methodology developed can provide rapid feedback on HOCIs that could be useful for infection prevention and control teams, and warrants further prospective evaluation. The integration of epidemiological and sequence data is important given the low mutation rate of SARS-CoV-2 and its variable incubation period.

Nosocomial transmission of SARS-CoV-2 presents a significant health risk to both vulnerable patients and to healthcare workers (HCWs) [1] [2] [3] [4] [5] . There is a variable incubation period, extending up to day 14 from exposure to the virus in symptomatic cases [6] . It is also known that transmission is possible from asymptomatic or presymptomatic carriers [7] [8] [9] [10] , complicating identification of hospital-acquisition among hospital onset COVID-19 infections (HOCIs) and tracing of likely sources of infection.

There is now substantial evidence from retrospective studies that genome sequencing of epidemic viruses, together with standard infection prevention and control (IPC) practice, better excludes nosocomial transmissions and better identifies routes of transmission than IPC investigation alone [11] [12] [13] . The development of rapid sequencing methods capable of generating pathogen genomes within 24-48 hours has recently created the potential for clinical IPC decisions to be informed by genetic data in near-real-time [14] . Although SARS-CoV-2 has a low mutation rate, estimated at around 2 changes per genome per month [15] , sufficient viral diversity does now exist for viral sequences to provide information regarding potential transmission clusters [16] . However, phylogenetic methods alone cannot reliably identify linked infections or the direction of transmission, and the need for clinical teams to gather additional patient data presents challenges to the timely interpretation of SARS-CoV-2 sequence data.

To overcome these barriers, we have developed a sequence reporting tool (SRT) that integrates genomic and epidemiological data from HOCIs to rapidly identify closely matched sequences within the hospital and assign a probability estimate for nosocomial infection. The output report is designed for prospective use to reduce the delay from SARS-CoV-2 sequencing to application of insights generated to IPC practice. As such, the probability model on which it is based has minimal computational requirements or need for local tuning and checking of model parameters. The work was conducted as part of the COVID-19 Genomics (COG) UK initiative, which sequences large numbers of SARS-CoV-2 viruses from hospitals and the community across the UK [17] . Here we describe the performance of the SRT using COG-UK sequence data for HOCI cases collected from Glasgow and Sheffield between February and May 2020 and explore how it may have provided additional useful information for IPC investigations.

The SRT methodology is applied to HOCI cases, defined here as inpatients with first positive SARS-CoV-2 test or symptom onset >48 hours after admission, who were not suspected of having COVID-19 at admission. The SRT algorithm returns probability estimates for healthcareassociated infection (HCAI) in each HOCI case, i.e. the probability that they acquired their infection post-admission within the hospital, with information provided on closely matching viral sequences from the ward location at sampling and wider hospital. Results for individual HOCIs are evaluated in relation to the IPC classification system recommended by PHE and other UK public health bodies, based on interval from admission to positive test: 3-7 days post admission = indeterminate HCAI; 8-14 days post admission = probable HCAI; >14 days post admission = definite HCAI [18] . We also applied the PHE definition of healthcare-associated COVID-19 outbreaks [18] (i.e. ≥ 2 cases associated with a specific ward, with at least one being a probable or definite HCAI) to ward-level data, and within each PHE-defined outbreak event we evaluated whether there was one or more clearly distinct genetic cluster. This was determined by consecutive linkage of each HOCI into clusters using a 2 SNP threshold (with HOCIs assigned to a genetic cluster if they were a sequence match to any member). Sequences with <90% genomic coverage were excluded from all analyses.

Research Ethics for COG-UK was granted by the PHE Research Ethics and Governance group as part of the emergency response to COVID-19 (24 April 2020, REF: R&D NR0195).

During the first wave of SARS-CoV-2, the MRC-University of Glasgow Centre for Virus Research collected residual clinical samples from SARS-CoV-2 infected individuals following diagnosis at the West of Scotland Specialist Virology Centre. Samples were triaged for rapid sequencing using Oxford Nanopore Technologies (ONT) for suspected healthcare related infections or Illumina sequencing in all other cases (details in supplementary Appendix).

Residual clinical samples from SARS-CoV-2 positive cases diagnosed at Sheffield Teaching Hospitals NHS Foundation Trust were sequenced at the University of Sheffield using ARTIC network protocol [19] and ONT. Throughout the epidemic, members of the IPC team were notified by the laboratory and by clinical teams of positive results and reviewed relevant areas to ensure optimisation of practice and appropriate management of patients. Electronic reports were created contemporaneously, including an assessment as to whether suspected linked cases were present based on ward level epidemiology. As part of SRT validation, these reports were accessed retrospectively by a study team member blind to the sequencing data and each included HOCI case was defined as being thought unlinked to other cases, a presumed index case in an outbreak or a presumed secondary case.

The sequence matching and probability score algorithm is run separately for each 'focus sequence' corresponding to a HOCI. We use associated metadata to assign other previously collected sequences to categories representing where the individual may be part of a SARS-COV-2 transmission network:

• Unit reference set: individual could be involved with transmission on same unit (ward/ICU etc) as focus sequence (look-back interval: 3 weeks)

• Institution reference set: individual could be involved with transmission in same institution/hospital as focus sequence (look-back interval: 3 weeks)

• Community reference set: individual could be involved with transmission outside of focus sequence institution (look-back interval: 6 weeks). It is possible for samples to be members of multiple reference sets. For example an outpatient may be involved in SARS-CoV-2 transmission at the institution they attended and/or in community transmission.

For each run of the algorithm, pairwise comparisons are conducted between the focus sequence and each sequence within the unit reference set, institution reference set and community reference set. A reference set sequence is considered a close match to the focus sequence if there is a maximum of two SNP differences between them. This choice was based on reported healthcare-associated outbreak events [14, 20] and the overall mutation rate of SARS-CoV-2 (details in supplementary Appendix).

We use an expression of Bayes theorem to estimate probabilities for post-admission infection of each focus case divided by exposure on the unit, within the rest of the institution and from visitors (if they were allowed). An estimate of the prior probability (P prior ) of post-admission infection for each focus case is modified to a posterior probability according to information provided by the sequence data. The algorithm is based on sound statistical principles, but involves heuristic approximations.

In symptomatic focus cases we base P prior on the time interval (t) from admission to date of symptom onset or first positive test (if date of symptom onset not recorded). We calculate P prior = F(t), where F() is the cumulative distribution function of incubation times [6] . The derivation of this method is given in the supplementary appendix.

In theory, it would be optimal to use all of the information in the exact sequences observed. However, with the goal of constructing a computationally simple algorithm, we base our calculations on the probability of observing a similar sequence (within 2 SNPs) to that actually observed for each focus case conditional on each potential infection source/location. We require estimates of the probability of observing a similar sequence to the focus sequence conditional on infection in the community, current unit/ward or elsewhere in the hospital/institution, or from a visitor. For the unit and hospital, we estimate this using the observed sequence match proportion (on pairwise comparison to the focus sequence) in the unit reference set and institution reference set, respectively. For community-or visitor-acquired infection we use a weighted proportion of matching sequences in the community reference set, with weightings determined by a calibration model that describes geographic clustering of similar sequences among community-acquired infections (described in supplementary appendix). The geographic weighting model was fitted separately for each study site using sequences strongly thought to represent community-acquired infection: all community-sampled sequences and patients presenting to the Emergency Department with COVID-19, excluding those recorded as being healthcare workers.

The analysis was conducted in R (v. 4.0.2, R Foundation, Vienna), using sequence processing and comparison functions from ape (v5.4) and geospatial functions in the PostcodesioR (v0.1.1) and gmt packages (v2,0). The algorithm has also been implemented as a standalone SRT for prospective use [21] within COV-GLUE [22] .

. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 15, 2020. ; https://doi.org/10.1101/2020.11.12.20230326 doi: medRxiv preprint

The Glasgow dataset included 1199 viral sequences (available as of 23 rd June 2020): 426 of which were derived from community sampling sites, 351 from patients presenting to the Emergency Department or hospital assessment locations such as acute medical units, 398 from hospital inpatients and 24 from outpatients. Limited data were available regarding the total number of HCWs testing positive and identification of their samples amongst those from the community-sampling sites, but 15 sequences were recorded as having been obtained from HCWs. The first positive test dates ranged from 3rd March to 27th May 2020. All consensus sequences had genomic coverage >90%.

We applied the SRT algorithm to data from three hospitals with the required metadata available, for which 128/246 inpatient cases with sequences were HOCIs. Two of these patients had been transferred from another hospital within 14 days prior to their positive test and were not processed as focus sequences. One patient for whom we were unable to determine the exact sampling location within the hospital was also excluded, leaving 125 HOCI cases for analysis. Population sequencing coverage was 536/1578 (34.0%) overall for patients at the three hospitals and 128/328 (39.0%) for HOCIs specifically ( Figure S1 ).

The Sheffield dataset included 1630 viral sequences with accompanying metadata (available as of 10 th October 2020): 714 were from inpatients, 117 were from outpatients and 799 were from HCWs. For the purpose of the retrospective evaluation, the 447/714 inpatient samples taken on date of admission were assumed to represent community-onset cases and used to calibrate the model. The first positive test dates ranged from 23rd February to 30th May 2020. One sequence with genome coverage <90% was dropped from further analysis (from an inpatient on date of admission). 201 of the inpatients were HOCIs. Population sequencing coverage was 714/977 (73.1%) overall for inpatients, 201/261 (77.0%) for HOCIs specifically and 799/962 (83.1%) for HCWs.

The SRT algorithm results in comparison to standard PHE classifications are summarised in Figure 1 and Table 1 . The majority of HOCI cases in Glasgow (78/125, 62.4%) and over a third of those in Sheffield (71/201, 35.3%) met the PHE definition of a definite HCAI and so are known to have acquired the virus post-admission irrespective of sequencing results. The probable HCAI cases formed the next largest group at each site. Overall, the SRT algorithm identified close sequence matches from the same ward for 66.4% of definite and 64.2% of probable HCAIs, indicating likely within-ward transmission (examples provided in Case Studies 1-3). When one or more close sequence matches was identified on the ward of the focus sequence, the SRT probability of infection on the ward was >0.5 in 185/189 cases (Figure 2 ). . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 15, 2020. ; https://doi.org/10.1101/2020.11.12.20230326 doi: medRxiv preprint For indeterminate HCAIs the SRT probability of HCAI was >0.5 in 33/82 (40.2%), and in 27/33 (81.8%) cases a close sequence match on the ward was present. Overall, 14/125 (11.2%) HOCIs in Glasgow and 175/201 (87.1%) in Sheffield had at least one close sequence match to a HCW sample, reflecting the much greater availability of sequences from HCWs in the Sheffield dataset.

In 16/244 (6.6%) cases that met the probable or definite HCAI definitions, there was no sequence match within the hospital; this is likely due to incomplete sequence data from SARS-CoV-2 hospitalised cases and staff (with population sequencing coverage <40% patients and very limited for staff from Glasgow and ≈ 75% of patients and staff in Sheffield) and the presence of undiagnosed carriers. To reflect this we designed the SRT report to return the following message in such situations "This is a probable/definite HCAI based on admission date, but we have not found genetic evidence of transmission within the hospital". There were 26 HOCIs in the Sheffield dataset for whom it was recorded that visitors were allowed on the ward at time of sampling. In three of these the estimated probability of infection from a visitor was between 0.4 and 0.5 (all had ≥ 18 days from admission to diagnosis and no close sequence matches on the ward).

Within the Sheffield dataset we identified six wards with two genetically distinct outbreak clusters (of two or more patients) and three wards with three distinct outbreaks (see Case Study 3). Standard IPC assessment had classified each as a single outbreak. We also identified 10 and 44 HOCIs in the Glasgow and Sheffield datasets, respectively, with no apparent genetic linkage to other HOCI cases on the ward but who met the PHE definition of inclusion within an outbreak event (Table 2 ). There were two HOCIs in the Sheffield dataset which showed a close sequence match to another case on the same ward with interval from admission to sample date ≤ 2 days.

. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 15, 2020. ; https://doi.org/10.1101/2020. 11.12.20230326 doi: medRxiv preprint Comparison to local IPC conclusions in Sheffield Detailed contemporaneous notes collected by IPC teams in Sheffield classified 18/201 HOCIs as the index case in outbreaks. IPC staff defined an index case as the first detected in an environment regardless of prior inpatient stay and, correspondingly, of these 14/18 were the first sequence on their ward and one was the second (the first 1 day earlier from a different bay on the ward was also recorded as an index case, and IPC staff deemed a ward outbreak with unclear index or possibly 2 index cases). Of the 18 index cases 11 showed at least one subsequent close sequence match on the same ward (the 2 index cases on a single ward were not genetically similar, and for 1/18 there were no subsequent sequences from the ward). The median SRT probability of HCAI was 0.70 (IQR 0.22-1.00, range 0.04-1, >0.5 in 12/18).

A further 144/201 HOCIs were classified as being part of local outbreaks, and among these the median SRT probability of HCAI was 0.98 (IQR 0.89-1.00; range 0.02-1.00; >0.5 in 129/144) with one or more close sequence match on the same ward in 104/144. The remaining 39/201 HOCIs, including 10 that were not recorded as HOCIs at the time, were classified by the IPC teams as not being part of local outbreaks). Among these the median SRT probability of HCAI was 0.74 (IQR 0.23-0.99, range 0.02-1.00; >0.5 in 23/39), with one or more close sequence matches on the same ward in 7/39. Figure 3 shows a phylogenetic tree of eight HOCIs within a single ward at a Glasgow hospital (Hospital 5, Unit 93), alongside associated meta-data and SRT probability outputs. The first HOCI detected (UID0032) was transferred from another hospital within the previous 2 weeks and so SRT output was not generated. All subsequent HOCIs return close sequence matches to at least one prior case on the ward, leading to SRT probability estimates of ward-acquired infection >0.9, even for UID0017 (an indeterminate HCAI). The phylogenetic tree indicates UID0032 has a SNP lacked by most of the cases identified on the ward, and therefore did not seed all of the cases in the outbreak cluster. Also shown on the tree is a single HOCI from a different ward in the same hospital (UID0025); this individual was an indeterminate HCAI, but a higher proportion of similar viral sequences within the hospital in comparison to their local community led to a SRT result of probable hospital-acquired infection.

Case Study 2 Figure 4 shows a phylogenetic tree indicating complex transmission networks across multiple hospitals in the Glasgow area (with SRT outputs for Hospitals 2 and 4). A monophyletic cluster of HOCIs can be seen in Hospital 2 Unit 48, with the first detected case identified by the SRT as a hospital-acquired and the rest unit-acquired infections. A paraphyletic group of HOCIs was detected in Hospital 4 Unit 69. Patient 1 (UID0042) was screened for COVID in Unit 69 on 14.04.20 after developing a cough and oxygen requirement. The patient was moved from the nightingale area to a single room on the ward on 14.04.20 and was confirmed positive on 15.04.20.

On 20.04.20 a second patient on Unit 69 (not sequenced) was screened after developing a cough and pyrexia and confirmed positive on 21.04.2020. The patient was in a single room at the time of symptom onset, however they had been in the main nightingale ward opposite patient 1 for 5 days. At this point 13 asymptomatic contacts in Unit 69 were screened, and 8 (UID0043, UID0073, UID0041, UID0095, UID0116, UID0094, UID0083, UID0121) were positive. These cases are all identified as hospital-acquired or unit-acquired infections and can be grouped into a genetically similar cluster with a maximum pairwise distance of 2 SNPs between each member and its nearest neighbour. However, this cluster clearly represents multiple introductions of SARS-CoV-2 onto the ward.

Case Study 3 Figure 5 shows phylogenetic trees relating to three distinct viral lineages identified on a single ward in the Sheffield dataset (classified by contemporaneous IPC investigation as a single outbreak). Two of these lineages also include sequences from inpatients sampled from other wards within the same hospital. Detailed ward movement data highlighted additional possible links between patients in the B.2.1 cluster. Both UID0149 and UID0157 were present at LOC0111 prior to their sample dates.

. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 15, 2020. ; https://doi.org/10.1101/2020.11.12.20230326 doi: medRxiv preprint

We have developed a novel approach for identification and investigation of hospital-acquired SARS-CoV-2 infections combining epidemiological and sequencing data, designed to provide rapid and concise feedback to IPC teams working to prevent nosocomial transmission. Through retrospective application to clinical datasets, we have demonstrated that the methodology is able to provide confirmatory evidence for most PHE-defined definite and probable HCAIs and provide further information regarding indeterminate HCAIs. Thus the SRT may allow IPC teams to optimise their use of resources, concentrating audit, review and educational interventions on areas with likely nosocomial acquisition events.

While the SRT is not likely to change IPC conclusions in cases meeting the definition of 'definite' or 'probable' HCAI based on interval from admission to symptom onset, in 91% of cases it did identify patients in the same ward or elsewhere in the hospital who could plausibly be linked to the HOCI within a single outbreak event. Those definite and probable HOCIs without close sequence matches are likely to reflect transmission from sources within the hospital that have either not been diagnosed or who were diagnosed without viral sequencing. In such cases it is impossible to calculate a probability of transmission and the SRT will simply state that no sequence matches were found within the hospital.

For cases meeting the definition of 'indeterminate healthcare associated', the probability scores returned would be useful for IPC teams. These probabilities are dependent on comparison to sequences from cases of community-acquired infection obtained either from direct community sampling or from patients sampled at admission. The Sheffield dataset was lacking the former data source, but the SRT nonetheless classified a similar proportion of 'indeterminate healthcare associated' HOCIs as community-acquired infections to that found in the Glasgow dataset (approximately 60%).

Current PHE guidelines define healthcare-associated COVID-19 outbreaks as two or more cases associated with a specific setting (e.g. ward), with at least one case having illness onset after 8 days of admission [18] . However, the guidelines note that "investigations of healthcare associated SARS-CoV-2 infection should also take into account COVID-19 cases categorised as 'indeterminate healthcare associated' (i.e. onset 3-7 days after admission), for which our SRT output would be useful. In most HOCIs meeting this definition of inclusion within an outbreak event, we found evidence of clusters of similar viral sequences located on the ward concerned, and the SRT results were in line with available local IPC classifications in the majority of cases. However, a substantial minority (54/279) of HOCIs although assumed to be part of a ward outbreak, were, in fact, isolated cases for which the sequencing data refuted genetic linkage to other sequences from the ward. The SRT also provided evidence of wards where IPC-defined outbreak events comprised two or three clearly distinct viral lineages (Case Report 3).

The retrospective datasets analysed in this study represent the first few months of the COVID-19 epidemic in the UK, and nosocomial transmission of the virus in the UK during this period has previously been reported at multiple sites [14, 23, 24] . HCWs were at increased risk of infection . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 15, 2020. ;  and adverse health outcomes [1, 2, 4, 5, 25] and could have been important drivers of nosocomial transmission [8] . Data were limited for Glasgow but the Sheffield dataset contained a large number of sequences obtained from HCWs, with population sequencing coverage for this group >80%, and there was a close sequence match to at least one HCW observed for 87% of HOCIs. Our analysis has not evaluated direction of transmission to or from HCWs, but they were clearly linked into transmission networks within the hospital. A limitation of the current SRT approach and of the retrospective data available is that they do not include detailed information regarding work locations for HCWs. However, prospective use of the SRT would allow IPC teams to investigate linkage from a HOCI to any HCWs flagged as having a close sequence match.

While a phylogenetic approach is useful in excluding direct transmission between cases, it can be more problematic to confirm transmission source [26] . Phylogenetic models can evaluate the full genetic information provided by viral sequence data, but there are challenges in incorporating and summarising associated patient meta-data in a timely fashion [27] . There will be cases in which phylogenetic analysis would provide information beyond that returned by the SRT. However, fully integrated epidemiological and phylogenetic analysis of hospital outbreaks is resource-intensive, presenting challenges in delivering the rapid turnaround and scale-up required to provide clear feedback to hospital IPC teams outside of research-intensive settings.

Comparison of SRT output to phylogenetic trees in a number of test cases suggested that some clusters of genetically similar cases identified within a specific ward likely represented more than one transmission event onto the ward from similar viral lineages circulating within the healthcare system. Whilst monophyletic clusters associated with a single location are easier to interpret, we consider the presence of viruses within a ward or hospital that are genetically similar to a HOCI as evidence for nosocomial infection even when they are not plausible transmission sources themselves, given the potential for asymptomatic transmission [7] [8] [9] [10] and complex transmission networks [14] .

The SRT uses a number of heuristic approximations in order to provide an integrated summary of epidemiological and sequence data. However, this choice is associated with the limitation that it does not provide a full probabilistic model of potential transmission networks. Further development of the SRT would also aim to more fully incorporate patient movement data and shift locations for HCWs.

Our novel approach to the investigation of HOCIs has shown promising characteristics on retrospective application to two clinical datasets. The SRT described allows rapid feedback on HOCIs that integrates epidemiological and sequencing data to generate a simplified report at the time that sequence data become available. Its prospective use will be evaluated in a multicentre trial in late 2020 and early 2021. The methodology has been developed for hospital inpatients, but the principles may also be applicable to other settings.

Acknowledgements . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 15, 2020. ; Days from first to last case in outbreak, median (IQR, range) 6 (4-9, 0-15) 4 (2-8, 0-17)

n wards with more than one distinct outbreak event 0 9* *In three wards there were three genetically distinct outbreak events.

. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this this version posted November 15, 2020. ; Figure 2 Plot of the posterior probabilities of healthcare-associated infection (HCAI) estimated using the sequence reporting tool algorithm from a source on the current ward versus a source elsewhere in the hospital for (a) Glasgow and (b) Sheffield HOCIs grouped by standard PHE classification. In cases where there are no close sequence matches in the dataset (including among community cases), the results returned are based solely on the priors and the metadata; this explains the fact that there are some cases with estimated posterior probability of infection on the ward greater than 0.5 for whom there were no sequence matches on the ward. a;

. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this this version posted November 15, 2020. ; . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this this version posted November 15, 2020. ; https://doi.org/10.1101/2020.11.12.20230326 doi: medRxiv preprint or . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this this version posted November 15, 2020. ; https://doi.org/10.1101/2020.11.12.20230326 doi: medRxiv preprint

Glasgow Sequencing with ONT followed the protocols developed by the ARTIC network (v1 and v2) https://artic.network/ncov-2019. The reads were aligned to the reference strain (MN908947) using minimap2 (https://doi.org/10.1093/bioinformatics/bty191) and denoised using nanopolish (https://www.nature.com/articles/nmeth.3444) prior to primer trimming and consensus calling with iVar using a minimum depth of 20 reads (https://doi.org/10.1186/s13059-018-1618-7). Sequencing with Illumina also used the ARTIC network protocol for amplicon generation but was followed by a DNA KAPA library preparation kit (Roche) and indexing with NEBNext multiplex oligos (NEB) using 7 PCR cycles. Libraries were pooled and loaded on a MiSeqV2 cartridge. Illumina reads were processed with the PrimalAlign pipeline (https://github.com/rjorton/PrimalAlign). Briefly, reads were trimmed using trim_galore (https://www.bioinformatics.babraham.ac.uk/projects/trim_galore/) aligned to the reference using BWA (10.1093/bioinformatics/btp698). Then, amplicon primers were removed and the consensus called with a read depth of 10 using iVar (https://doi.org/10.1186/s13059-018-1618-7). Metadata associated with each sample was collated in a redcap database (https://www.project-redcap.org/).

Sequencing with ONT followed the protocols developed by the ARTIC network (v1 and v2) https://artic.network/ncov-2019. Following base calling, data were demultiplexed using ONT Guppy using a high accuracy model. Reads were filtered based on quality and length (400 to 700bp), then mapped to the Wuhan reference genome and primer sites trimmed. Reads were then downsampled to 200x coverage in each direction. Variants were called using nanopolish (https://github.com/jts/nanopolish) and used to determine changes from the reference. Consensus sequences were constructed using reference and variants called.

Data sources for algorithm There are two potential sources of data for the HOCI classification algorithm. Firstly, there are institution-sampled sequences: these include all viral sequences from samples obtained within the institution/hospital. These sequences are linked to meta-data providing basic information regarding the patient concerned and details of the sample from which the sequence was obtained. Secondly, there are community-sampled sequences: these include all relevant sequences obtained from samples from testing within the local community. These sequences are associated with a more limited set of linked meta-data describing date of sample, residential outer postcode of subject and place of work if they are recorded as being a HCW.

. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 15, 2020. ; https://doi.org/10.1101/2020.11.12.20230326 doi: medRxiv preprint Unit reference set This data set comprises all institution-sampled sequences sampled on or ≤ 3 weeks prior to (or ≤ 2 days after for the prospective version of the SRT) the sample date of the focus sequence and for which both the institution and the unit is the same as that for the focus sequence.

This data set comprises firstly all institution-sampled sequences from HCWs, outpatients and inpatients diagnosed >48 h after admission for which the institution matches that of the focus sequence sampled on or ≤ 3 weeks prior to (or ≤ 2 days after for the prospective version of the SRT) the sample date of the focus sequence and for which the unit is either not the same as that for the focus sequence or is missing. Secondly, the data set includes all institution-sampled sequences from A&E patients or inpatients diagnosed ≤ 2 days after admission for which the institutionID matches that of the focus sequence sampled between (inclusively) 3 weeks and 3 days prior to the sample date of the focus sequence and for which the unit is either not the same as that for the focus sequence or is missing. Thirdly, this data set also includes the subset of community-sampled sequences of healthcare workers at the same institution as the focus sequence.

This data set comprises firstly all community-sampled sequences sampled on or ≤ 6 weeks prior to (or ≤ 2 days after for the prospective version of the SRT) the sample date of the focus sequence. This data set also includes institution-sampled sequences sampled on or ≤ 6 weeks prior to (or ≤ 2 days after for the prospective version of the SRT) the sample date of the focus sequence from all non-inpatient samples, and those inpatients for whom sample date and symptom onset date (if recorded) are both ≤ 2 days after the admission date.

Note that some institution-sampled sequences will contribute to both the community reference set and either the unit reference set or the institution reference set (e.g. outpatients sampled within 3 weeks prior to the focus sequence would be included in both the community reference set and the institution reference set). HCWs recorded among the community-sampled sequences within ≤ 3 weeks prior to the sample date of the focus sequence will also be included in both the community reference set and the institution reference set if their workplace matches the institution of the focus sequence.

ܲ ‫כ‬ ܲ ௨ ‫כ‬ ܲ ሺ ‫ݏ‬ ݁ ‫ݍ‬ േ 2 ܵ ܰ ܲ ‫ݏ‬ | ܷ ‫ܫ‬ ሻ ܲ ‫כ‬ ܲ ௨ ‫כ‬ ܲ ሺ ‫ݏ‬ ݁ ‫ݍ‬ േ 2 ܵ ܰ ܲ ‫ݏ‬ | ܷ ‫ܫ‬ ሻ ܲ ‫כ‬ ܲ ௩ ‫כ‬ ܲ ሺ ‫ݏ‬ ݁ ‫ݍ‬ േ 2 ܵ ܰ ܲ ‫ݏ‬ | ܸ ‫ܫ‬ ሻ ܲ ‫כ‬ ሺ 1 െ ܲ ௨ െ ܲ ௩ ሻ ‫כ‬ ܲ ሺ ‫ݏ‬ ݁ ‫ݍ‬ േ 2 ܵ ܰ ܲ ‫ݏ‬ | ‫ܫ‬ ‫ܫ‬ ሻ ሺ 1 െ ܲ ሻ ‫כ‬ ܲ ሺ ‫ݏ‬ ݁ ‫ݍ‬ േ 2 ܵ ܰ ܲ ‫ݏ‬ | ‫ܥ‬ ‫ܫ‬ ሻ

ܲ ‫כ‬ ሺ 1 െ ܲ ௨ െ ܲ ௩ ሻ ‫כ‬ ܲ ሺ ‫ݏ‬ ݁ ‫ݍ‬ േ 2 ܵ ܰ ܲ ‫ݏ‬ | ‫ܫ‬ ‫ܫ‬ ሻ ܲ ‫כ‬ ܲ ௨ ‫כ‬ ܲ ሺ ‫ݏ‬ ݁ ‫ݍ‬ േ 2 ܵ ܰ ܲ ‫ݏ‬ | ܷ ‫ܫ‬ ሻ ܲ ‫כ‬ ܲ ௩ ‫כ‬ ܲ ሺ ‫ݏ‬ ݁ ‫ݍ‬ േ 2 ܵ ܰ ܲ ‫ݏ‬ | ܸ ‫ܫ‬ ሻ ܲ ‫כ‬ ሺ 1 െ ܲ ௨ െ ܲ ௩ ሻ ‫כ‬ ܲ ሺ ‫ݏ‬ ݁ ‫ݍ‬ േ 2 ܵ ܰ ܲ ‫ݏ‬ | ‫ܫ‬ ‫ܫ‬ ሻ ሺ 1 െ ܲ ሻ ‫כ‬ ܲ ሺ ‫ݏ‬ ݁ ‫ݍ‬ േ 2 ܵ ܰ ܲ ‫ݏ‬ | ‫ܥ‬ ‫ܫ‬ ሻ

. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 15, 2020. ; https://doi.org/10.1101/2020.11.12.20230326 doi: medRxiv preprint

When there is a close sequence match found in any of the defined reference sets, the posterior probability estimates for UI, II, VI and CI will always sum to 1. However, when there is no close sequence match in any of the reference sets the posterior probability calculations are not valid and the algorithm will return the prior probabilities for each potential source/location of infection.

Further details regarding sequence matching process

The ±2 SNP threshold for a close sequence match was initially based on reports of healthcareassociated outbreak events for which this was the maximum pairwise difference within clusters . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 15, 2020. ; https://doi.org/10.1101/2020.11.12.20230326 doi: medRxiv preprint • definite HCAI: P prior ≥ 0.995

For asymptomatic focus cases, we define our prior on the basis that some proportion of the cases detected will never become symptomatic (P a ) with the remainder going on to develop symptoms within the next few days (1-P a ). We then define our prior probability of post-admission infection in these cases as:

where t is the interval from admissionDate to sampleDate and c is a constant reflecting the average interval within which we expect symptoms to appear (among those cases in which they do). P a is set at 0.4 based on the findings of a published review article (Oran and Topol: doi.org/10.7326/M20-3012), and c is set to 3 based on a combination of expert opinion of the study PIs, the known distribution of time from infection to symptom onset and expert experience of asymptomatic screening.

The model requires prior values for the probability of UI and VI given post-admission infection: P u and P v , respectively. However, in specifying the model we define P u ' as the probability of UI given post-admission infection when there are no visitors allowed on the ward, in which case the probability of VI is zero and P v '=0. If visitors are allowed on the ward for the focus case, then we set P u = P u '×(1-P v ).

Based on expert opinion of the clinical co-coauthors, P u ' is set to different values according to the unit/ward type of the focus sequence with single bed wards having a lower prior probability of unit post-admission infection than bay wards: 0.5 for single bed wards and 0.7 for bay wards. We assumed a P v of 0.2. The P u values (when visitors are allowed) are therefore: 0.4 for single bed wards and 0.56 for bay wards. The largest of the three Glasgow hospitals included comprises single-room wards, whilst the other two and the Sheffield site comprise bay wards.

If we assume a uniform individual-level hazard (λ) of infection from 1st February 2020 (t 0 ), whether in hospital or not, then the probability density function (PDF) of infection at time t inf from this date is: λ e^(-λt inf ). The PDF of infection at time t inf conditional on this occurring at any point prior to the date of symptom onset (t onset ) is: (λe^(-λt inf )) / (1-e^(-λt pos )), which is approximately 1/t onset for small λ (taking the limit as λ ->0). For HOCI cases, we are interested in whether t inf occurred before or after the time of admission to hospital (t adm ). Also considering the evidence provided by the known incubation time of the disease (PDF f and CDF F), we integrate over the range of possible infection dates:

Geographic weighting function The weight of each sequence within the community reference set is determined by geographic distance from the residential outer postcode of the focus case, using a function of the form:

β takes a value between 0 and 1, and τ >0. These parameters are set based on calibration to the available community reference set at each site. The rationale for this weighting is that there is likely to be geographic clustering of viral lineages, and so newly observed community transmissions of SARS-CoV-2 are more likely to show genetic similarity to past sequences from the local area of that individual's home than to past sequences from regions that are further away. If postcode is missing for a case in the community reference set, then distance to the focus sequence is set to 100 km.

The statistical model for geographic weighting is fitted separately for each study site using sequences which are strongly thought to represent community-acquired infection: all community-sampled sequences and patients presenting to A&E with COVID-19, excluding those who are recorded as being healthcare workers or who do not have an available valid outer postcode. We will refer to these sequences as the 'calibration set'.

A statistical model is constructed to find the optimal values of β and τ to maximise the estimated probability (P sim:i ) of a newly observed community-acquired case having a similar sequence (±2SNPs) to that observed for each sequence in the calibration set. The estimated probability in each case within the calibration set is calculated as a weighted sum of 'close match' indicator variables for all other sequences in the calibration set sample from 6 weeks prior up until the sample date of that case, with the weighting function defined in terms of geographic distance between residential outer postcodes and the β and τ parameters as described for the community reference set.

An overall log-likelihood function is defined using a Bernoulli distribution for each of the n sequences within the calibration set: We assume that the probability of a sequence match conditional on infection from visitor on unit/ward can be calculated using the same weighting scheme as for the probability of a sequence match conditional on community-acquired infection (i.e. P(seq±2 SNPs|CI)==P(seq±2 SNPs|VI)).

. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 15, 2020. ; https://doi.org/10.1101/2020. 11.12.20230326 doi: medRxiv preprint Additional matching on ward location history There is the potential for the algorithm described to return large numbers of close sequences matches with the hospital as a whole, which may make it difficult for IPC teams to use the output to direct their investigations when there are no potential sources of infection identified on the same ward as the focus case. We propose a location matching procedure in order to highlight the most relevant sequence matches for further investigation. This process does not currently form part of the statistical model, meaning that it can be treated as optional functionality for the SRT in the COG-UK HOCI study, and we have restricted the input data to a simplified format in order to minimise data management requirements.

For each inpatient sample in the input meta-data for the algorithm, we specify a single string variable comprising the concatenated names of any ward locations in the ≤ 14 days prior to the sample date and a separate string variable with any ward locations in the ≤ 14 days after the sample date. For each focus case submitted to the algorithm, output is flagged if there is any match identified between the wards listed in each of these fields or the ward at time of sampling for a close sequence match in comparison to the prior and current ward locations for the focus sequence (excluding those cases were there is already matching ward location at time of sampling for each).

Phylogenies were produced by the grapevine pipeline (https://github.com/COG-UK/grapevine) as part of the COG-UK Consortium (https://www.cogconsortium.uk). Briefly, sequences from GISAID and those produced as part of the COG-UK Consortium are independently quality controlled and aligned to the Wuhan reference using minimap2 (https://doi.org/10.1093/bioinformatics/bty191). The two alignments are then combined, the homoplasy at site 11083 is masked and the tree is reconstructed using FastTreeMP (http://www.microbesonline.org/fasttree/). For each of the hospitals of interest, the tree is pruned to keep sequences from Scotland or Yorkshire (as relevant) and by date excluding sequences subsequent to the last "focus" patient sample date on the ward. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 15, 2020. ; https://doi.org/10.1101/2020.11.12.20230326 doi: medRxiv preprint Details of SRT report format

The SRT system for prospective use needs to provide useful and appropriate feedback in both low incidence and high incidence settings for new HOCI cases. This is planned through the generation of a concise one-page PDF summary report for each focus sequence. This summary report will contain key focus sequence meta-data, information regarding the estimated probabilities for infection source and details of up to ten close sequence matches identified within the same unit/ward and/or elsewhere in the hospital.

The sequence matching and probability score algorithm generates probability estimates for the source of infection for the focus patient being from the current unit/ward, from elsewhere in the hospital, from the community (pre-admission) or from a visitor. These probability estimates always sum to 1. In the summary report, probability estimates for each source of infection are categorised using the following levels: For clarity of presentation and communication, probability categories will not always be displayed in the summary report for all four potential sources of infection (i.e. ward/unit, elsewhere in hospital, visitor, or community). Special handling rules for specific situations are described below.

Close sequence matches within the same unit and/or hospital The maximum number of close sequence matches that can be listed on the one-page summary report is 10 (for the combined sum of unit-level and institution-level matches). If the number of ward-level matches is n>5 and the total number of close sequence matches is N>10, then the number of ward-level matches is truncated at 5+max((5-(N-n)),0). If there are over ten close sequence matches in total, then the following message is displayed "Over 10 close matches; see detailed report for further information".

Within the each the sets of unit-level and institution-level close sequence matches, ordering and priority for inclusion within the available slots is determined by the following set of criteria (in decreasing order of importance):

1. Number of SNPs relative to Wuhan strain present in comparison sequence but absent in focus sequence (fewer = higher priority) 2. Number of SNPs relative to Wuhan strain present in focus sequence but absent in comparison sequence (fewer = higher priority) 3. Whether comparison sequence is from a HCW (HCWs listed first) 4. HCAI status of comparison sequence (priority order: definite, probable, indeterminate, otherwise)

. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 15, 2020. ; https://doi.org/10.1101/2020.11.12.20230326 doi: medRxiv preprint 5. Samples from the past before samples in future 6. Samples from within the two weeks prior to focus sequence sample date before others 7. Number of units overlapping with focus sample's units

No close sequence matches on unit/ward If there are no close sequence matches to the focus sequence on their current unit/ward, then no probability category is reported for this potential infection source (the algorithm returns a zero probability in such cases, which could be misleading given uncertainty over screening and sequencing coverage). The message "No matches from within unit" is displayed. The probability score category for infection from elsewhere in the hospital is provided in such cases.

If there are no close sequence matches to the focus sequence elsewhere in the hospital, then no probability category is reported for this potential infection source. The message "No matches elsewhere in hospital" is displayed.

If the estimated probability of community-acquired infection from the algorithm is >50%, but the interval from admission to symptom onset (if recorded) or sample date is ≥ 8 days, then the following message is displayed in place of the estimated probability of community-acquired infection "This is a probable/definite HCAI based on admission date, but we have not found genetic evidence of transmission within the hospital".

If the posterior probability of unit-acquired infection and the posterior probability of infection from a source elsewhere in the hospital are each estimated to be <50%, but the sum of these two posterior probabilities is ≥ 50%, then the following message is displayed "Overall, this is a probable unit-or institution-acquired infection with source unclear".

The timeline graph provides a visual representation of available sequences from the same unit/ward and the same institution/hospital as the focus sequence in the period from 3 weeks prior to their sample date to 1 week after. The key indicates which sequences are close matches to the focus sequence, and the numbering corresponds to that in the tabular summary of most relevant close sequence matches.

. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 15, 2020. ;

Sequencing coverage in Glasgow dataset Figure S1 Proportion of cases sequenced in Greater Glasgow and Clyde Health Board between 1 March and 27 th May (with sequence available as of 23 June 2020) by location of test (A). Also displayed are the proportion of sequenced cases in the three focus hospitals subdivided by assessment and inpatient locations (B), and the proportion of HOCI cases sequenced at these hospitals (C).

. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 15, 2020. ; https://doi.org/10.1101/2020.11.12.20230326 doi: medRxiv preprint . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 15, 2020. ; https://doi.org/10.1101/2020.11.12.20230326 doi: medRxiv preprint Home residence locations and geographic model parameters . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 15, 2020. ; https://doi.org/10.1101/2020.11.12.20230326 doi: medRxiv preprint SNP distance distributions For the Glasgow sequence dataset as a whole the median pairwise SNP difference among all sequences was 9, and there were 1.3%, 3.4%, 6.4% and 10.1% of pairwise comparisons with 0, ≤ 1, ≤ 2 and ≤ 3 SNP differences, respectively. For the Sheffield dataset as a whole the median pairwise SNP difference among all sequences was 8, and there were 1.2%, 3.3%, 6.5% and 10.8% of pairwise comparisons with 0, is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) preprint

The copyright holder for this this version posted November 15, 2020. ; https://doi.org/10.1101/2020.11.12.20230326 doi: medRxiv preprint

Deaths in healthcare workers due to COVID-19: the need for robust data and analysis

Clinical Characteristics of 138 Hospitalized Patients With 2019 Novel Coronavirus-Infected Pneumonia in Wuhan, China

What settings have been linked to SARS-CoV-2 transmission clusters?

Scoping Report on Hospital and Health Care Acquisition of COVID-19 and its Control

Risk of hospital admission with coronavirus disease 2019 in healthcare workers and their households: nationwide linkage cohort study

The Incubation Period of Coronavirus Disease 2019 (COVID-19) From Publicly Reported Confirmed Cases: Estimation and Application

Temporal dynamics in viral shedding and transmissibility of COVID-19

Screening of healthcare workers for SARS-CoV-2 highlights the role of asymptomatic carriage in COVID-19 transmission

Prevalence of Asymptomatic SARS-CoV-2 Infection

Wholegenome Sequencing to Track Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) Transmission in Nosocomial Outbreaks

Norovirus Transmission Dynamics in a Pediatric Hospital Using Full Genome Sequences

Use of Whole-Genome Sequencing of Adenovirus in Immunocompromised Pediatric Patients to Identify Nosocomial Transmission and Mixed-Genotype Infection

Whole-genome Sequencing Provides Data for Stratifying Infection Prevention and Control Management of Nosocomial Influenza A

Rapid implementation of SARS-CoV-2 sequencing to investigate cases of health-care associated COVID-19: a prospective genomic surveillance study

Coast-to-Coast Spread of SARS-CoV-2 during the Early Epidemic in the United States

Emergence of genomic diversity and recurrent mutations in SARS-CoV-2. Infection

The COVID-19 Genomics UK (COG-UK) consortium. An integrated national scale SARS-CoV-2 genomic surveillance network

Professor Alison Holmes MD 44, 45 , Dr Robin Howe 33

Dr Rachael Stanley PhD 52 , Claire Stuart 52 and Dr Elizabeth Wastenge MD 20 . Sequencing and analysis: Dr Erwan Acheson PhD 6 , Safiah Afifi BSc 36 , Dr Elias Allara MD PhD 2, 3 , Dr Roberto Amato 1 , Dr Adrienn Angyal PhD 38 , Dr Elihu Aranday-Cortes PhD/DVM 21

Dr Natalie Groves 7 ,Dr James W Harrison PhD 49

Dr Samantha McGuigan 1 , Liz McMinn 1 , Dr Lamia Mestek-Boukhibar D.Phil 41