Antibiotic review kit for hospitals (ARK-Hospital): a stepped-wedge cluster-randomised controlled trial

mortality within 30 days of admission (patient level, non-inferiority margin of 5%). Outcomes were assessed in the modified intention-to-treat population (ie, excluding sites that withdrew before implementation). Intervention effects were assessed by use of interrupted time series analyses within each site, estimating overall effects through random-effects meta-analysis, with heterogeneity across prespecified potential modifiers assessed by use of meta-regression. This trial is completed and is registered with ISRCTN, ISRCTN12674243.


Introduction
The effect of antimicrobial resistance on global public health is similar to the effects of malaria and HIV, causing an estimated 4•95 million deaths in 2019. 1 Antimicrobial resistance places increased demands on health-care systems, with substantial economic conse quences. 2 Human antibiotic consumption is a major driver of antimicrobial resistance, 3 with increased use driving resistance at both a population level and an individual-patient level. 4lthough antibiotic use varies widely between and within health-care systems, no evidence exists that clinical outcomes are influenced by this wide variation (eg, between acute hospitals in England). 5ntimicrobial stewardship aims to minimise resistance selection by ensuring that antibiotics are prescribed only when clinically indicated and that narrow-spectrum agents are used whenever appropriate. 6However, translating research on antimicrobial stewardship into practice is hampered by the poor quality of evidence, particularly weaknesses in the intervention design process and the study designs used, which are usually underpowered, not experimental, and do not consider clinical outcomes. 7,8n primary care, restrictive strategies for antimicrobial stewardship, such as avoiding or delaying antibiotics in respiratory tract infection, can be safe and effective. 9,10By contrast, hospital antimicrobial stewardship interven tions that enable improved prescribing are more acceptable than restrictive strategies and can reduce overuse and length of stay without compromising mortality. 11The need to ensure that patients with serious bacterial infections are treated promptly before a diagnosis is confirmed means that ongoing review and revision of hospital antibiotic prescriptions is required to safely minimise unnecessary use.In England, the Department of Health's guidance Start Smart-Then Focus requires prescribers to review and revise antibiotic prescriptions every 48-72 h. 12 In the USA, the analogous term antibiotic timeouts is used, but revised Centers for Disease Control and Prevention guidance in 2019 prioritised pharmacist-led audit and feedback to prescribers, 13 in light of a non-interventional study that reported that prescriber-led reviews did not reduce overall consumption. 14fter introducing Start Smart-Then Focus in 2011, antibiotic consumption in English hospitals continued to rise year on year until the COVID-19 pandemic in 2020. 15his increase was despite financial incentives to reduce hospital prescribing, first through a Commissioning for Quality and Innovation framework in 2016-18 and then its incorporation into the National Health Service (NHS) standard contract for acute hospitals. 16Although high rates of prescription review were achieved, 17 most review decisions were to adjust rather than stop antibiotics.
The antibiotic review kit (ARK) programme for hospitals aimed to develop and evaluate a multifaceted behaviour change intervention to safely reduce antibiotic use in acute general medical inpatients.The ARK programme created a four-component interven tion to help prescribers to take appropriate decisions to stop or continue antibiotics at prescription review, comprising a novel prescribing decision aid, an online training tool supporting use of the decision aid, guidance for implementing audit and feedback, and a patient leaflet. 18,19n this study we report the immediate and sustained effect of the ARK intervention on antibiotic consumption at the hospital level and clinical outcomes at the patient level.

Study design
Following a feasibility evaluation at Brighton and Sussex University Hospitals NHS Trust, an acute hospital, 20 the ARK intervention was evaluated at eligible participating hospi tals across all UK nations by use of a stepped-wedge We searched PubMed, with no language or date restrictions, on Jan 31, 2022, for clinical studies that focused on improving antibiotic use for adults who were admitted to hospital using the terms "anti-bacterial agents therapeutic use" and "antibiotic stewardship".Among the 427 studies found, most were uncontrolled evaluations of different approaches to education, decision support, and feedback.These studies included one before-and-after study, which identified no effect of unsupported clinician-led prescription review on antibiotic use.Three small, hospital-level, cluster-randomised trials were identified.One trial evaluated different approaches to feedback, one compared different hospital specialties, and one reported that intense feedback was effective in reducing antibiotic use.All three trials were small and none considered clinical outcomes or sustainability.Research is needed to deliver effective interventions that are ready for implementation into clinical practice.This weak evidence base explains the differences that exist in national policy recommendations around clinician-led antibiotic prescription review for hospital antibiotic stewardship between, for example, the UK and the USA.

Added value of this study
We evaluated a multifaceted intervention to support clinicianled antibiotic prescription review (ie, the antibiotic review kit [ARK] intervention) and showed that ARK was effective in achieving safe sustained reductions in organisation-level antibiotic use among acute, general medical hospital admissions.Our findings deal with the uncertainty about whether clinician-led prescription review is an effective approach to antibiotic stewardship in hospital practice by being highly pragmatic, evaluating sustainability, and robustly exploring potential patient-level harms of this approach to reducing antibiotic use.Furthermore, the ARK-Hospital Programme delivers resources to support effective clinician-led prescription review ready for adoption into clinical practice.

Implications of all available evidence
The ARK intervention is safe and effective in reducing antibiotic use among adult acute, general medical hospital admissions.The tools used are now freely available for adoption into practice.Available evidence comes from research using paperbased prescribing and future research should establish how antibiotic prescription reviews should be built into electronic prescribing systems.
cluster-randomised controlled trial. 21A cluster design was essential to avoid contamination from health-care professionals moving between teams within a hospital.A stepped-wedge design was essential given the few UK secondary care organisations that could be randomised (appendix p 46).
Ethical approval was from the South Central Oxford C Research Ethics Committee (17/SC/0034) and the Confidentiality Advisory Group (17/CAG/0015) without individual patient consent because electronic health records were pseudonymised and no personal identifiable data was collected other than date of death.

Clusters and participants
The unit of observation was a hospital organisation offering services for non-elective medical admissions (appendix p 3). Sites were approached through professional networks and the Society for Acute Medicine.Eligible sites needed to admit adult (ie, aged >15 years) general or medical inpatients, have a local representative (known as a champion) who was willing to lead intervention implementation, and be able to provide the required study data.Since the intervention targeted prescribers on acute general medical wards and used electronic health records to ascertain patient-level outcomes, the study population was defined using the consultant specialty codes that were most often used to admit adult general medicine inpatients (appendix p 45). 21 Sites were asked to exclude patients who opted out of having their health records used for research purposes (appendix pp 3-4).The protocol is included in the appendix (p 156).

Randomisation and masking
Eligible sites were randomised by use of a computergenerated list by the trial statistician (ASW), including the pilot sites (one block of three sites) and main trial sites (six blocks of six sites and one block of seven sites), to an intervention implementation date.Implementation was staggered across sites in 1-2 week intervals, with breaks over the Christmas period and in August given the high rate of staff holidays or when the funder requested a pause on randomisation (appendix p 46).To avoid contamination, complete inform ation about the intervention and allocation sequence was concealed from the site until the point of randomisation, when sites were told that their randomised implementation date was 12 weeks in the future, ensuring that all sites had 12 weeks for implementation preparation.

Procedures
The intervention comprised a decision aid that was intended to be embedded in the hospital prescription process, prompting prescribers to clarify the level of diagnostic uncertainty at antibiotic initiation by classifying infection risk as possible or probable, and then either stopping the prescription if a clear indication for ongoing antibiotic treatment could not be established at 48-72 h review or finalising the prescription if a clear indication could be established; online training to motivate and support use of the decision aid; implementation guidance, includ ing audit and feedback tools; and a patient leaflet. 18,19he ARK tools are freely available through the British Society for Antimicrobial Chemotherapy.By supporting decisions to stop antibiotics at clinical review, the intervention aimed to safely reduce antibiotic use through reducing treatment duration, rather than by targeting the appropriateness of initial prescriptions.Fidelity of intervention implementa tion was assessed with eight predefined criteria by the study team at each individual site up to 16 weeks after implementation (table ).
Study data were collected from 24 months before implementation at the first main trial site until at least 14 months following implementation of the final site, to facilitate outcome assessment before and after implementation.Time periods for the co-primary outcomes are shown in the appendix (appendix p 46).
All outcomes were assessed using pseudonymised electronic health records from adult (age ≥16 years) acute general medical admissions (further details, including data cleaning, are shown in the appendix pp 6-9, 48), bulk antibiotic dispensing on the wards that implemented ARK, and C difficile test results.Date of death within 90 days of admission (in or out of hospital) was obtained by sites through linkage with national registries.Patient-level antibiotic data and laboratory results (ie, micro biology, haematology, biochemistry, and imaging tests) were provided by few sites, preventing further analysis of these data (appendix p 5). Uptake of the intervention was assessed through a process evaluation and prescription review audits.

Outcomes
The trial had two co-primary outcomes: antibiotic defined daily doses (DDDs) per adult acute general medical admission (superiority) and all-cause mortality within 30 days of admission (in or out of hospital; non-inferiority, relative margin 5% for an immediate step change associated with implementation, assuming a constant rate before and after implementation).Both outcomes were assessed by estimating the immediate effect and the sustained year-on-year effect.
Secondary antibiotic (superiority) outcomes were total antibiotic DDDs per acute general medical bed-day and DDDs per admission for specific antibiotic groups, including carbapenems, parenteral and oral administration, broad-spectrum and narrow-spectrum antibiotics, and the UK Health Security Agency's interpretations of Access, Watch, and Reserve from WHO's Essential Medicines List (appendix pp 11-12). 22Admissions, rather than bed-days, were used as the denominator in the primary analysis because bed-days can be influenced by non-medical reasons for prolonged hospital stays (eg, awaiting discharge to another place of care).Although the protocol specified patient-level antibiotic outcomes (ie, days on antibiotics, antibiotic-days per admission or bed-day, and antibiotic restart after discontinuation for >48 h), 21 only four sites had both the electronic prescribing systems and information tech nology resources required to provide these data, so these outcomes could not be analysed (figure 1).Sec ondary non-inferiority out comes were 90-day mortality, admission to an intensive care unit, length of stay, emergency hospital readmission (to any specialty) within 30 days of discharge, and Clostridioides difficile infection or colonisation within 90 days of admission.
DDDs per admission for piperacillin-tazobactam and quinolones and length of stay of 48 h or more were considered in exploratory analyses.

Statistical analysis
An interrupted time series analysis estimated the immediate effect of the intervention (ie, step change) and sustained effect on year-on-year trends after implementation versus before implementation within each site for the co-primary outcomes of antibiotic DDDs per adult acute general medical admission and all-cause mortality within 30 days of admission, by use of the randomised implementation date.All analyses used an intention-totreat approach that was modified to exclude seven sites that withdrew after randomisation but before implementation, from which no data were collected.Overall intervention effects were then estimated by use of random effects meta-analysis, using meta-regression to assess heterogeneity in effects across prespecified potential effect modifiers.As the intervention did not change after the pilot, following the approved protocol, the primary analysis included pilot and main trial sites to maximise power.With a minimum of 36 sites, the stepped-wedge cluster-randomised design had more than 85% power to exclude an immediate 5% relative increase in 30-day mortality and to detect a 15% relative reduction in antibiotic use associated with intervention implementa tion. 21he 5% relative non-inferiority margin for the 30-day mortality outcome means that to declare non-inferiority, the upper bound of 95% CI for the relative change in mortality had to not exceed a 5% increase.Further details are given in the appendix (appendix pp 9-10).
Monthly antibiotic DDDs per admission were modelled by use of negative binomial regression, and binary outcomes per admission were modelled by use of logistic regression.Length of stay (days) was modelled by use of subhazard regression, treating inpatient deaths as a competing risk and censoring at 90 days, using 0•1 days for those admitted and discharged on the same day, as was emergency 30-day readmission in a sensitivity analysis, with out-of-hospital deaths as the competing event.Due to low event rates (<4% in all sites), sensitivity analyses did not model ICU admission and C difficile infection or colonisation by use of subhazard regression.Length of stay of 48 h or longer was considered in an exploratory analysis by use of logit models and exploratory analyses for DDDs per admission for piperacillin-tazobactam and quinolones were modelled by use of negative binomial regression.All clinical outcomes were modelled by use of a robust variance adjustment by patient.
Since the COVID-19 pandemic profoundly affected both primary and secondary outcomes, models included a binary indicator for March-June, 2020, for antibiotic outcomes (measured monthly) and March 1-June 30, 2020, for patient-level clinical outcomes, unless otherwise noted.Sensitivity analyses excluded admissions after March 1, 2020, including 12 sites with less than 12 months of data after imple mentation as a result.Antibiotic models additionally adjusted for seasonal effects by including month of year as a sin() + cos() function to ensure smooth risk changes year to year.Non-antibiotic models also adjusted for individual admission-level covariates, regardless of statistical significance (based on the findings of Walker and colleagues 23 ): sex, age, immuno suppression, deprivation percentile, Charlson comor bidity index and its interaction with age, admission method, admission source, admission specialty, patient classifica tion, admis sion day of the week (ie, weekend vs weekday), admission day of year and time of day (both modelled as a sin() + cos() function, with an interaction between time of day and day of week), and number of overnight admissions and any previous overnight complex (ie, >1 consultant episode, excluding episodes in the emer gency department and rehabilitation) admission in the past year.Ethnicity was missing for a median of 8•8% admissions (IQR 4•5-18•4) per site so was not adjusted for.Further details are shown in the appendix (pp 4-5, 21-28).
All analyses used Stata/MP version 17.0.The data monitoring committee reviewed outcome data three times during the trial, using a Haybittle-Peto statistical rule for early stopping.This trial is registered with ISRCTN, number ISRCTN12674243.

Role of the funding source
The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report.

Results
58 UK acute hospital organisations expressed an interest in participating, of which 46 sites agreed to join the pilot or main trial.Three pilot sites implemented the intervention between Sept 25-Nov 20, 2017; 43 further sites were randomised to implement between Feb 12, 2018, and July 1, 2019, of which seven withdrew before imple mentation and were excluded from analyses as no data were collected (figure 1).39 sites were included in the analysis of 30-day mortality and 38 sites were included in the analysis of total antibiotic DDDs per adult acute general medical admission because antibiotic data were not available from one site (figure 1, appendix p 29).
13 sites were classed as large (ie, >850 beds available, median 991), 14 medium (ie, 551-850 beds, median 670), and 12 small (ie, ≤550 beds, median 487; table).Sites were distributed across the UK, with the largest number in the south of England.Most champions were microbiologists.At imple mentation, prescribing was paper based at most sites.21 (54%) of 39 sites implemented the decision aid with a hard stop to the initial prescription unless revised by 72 h, nine sites (23%) implemented as a soft stop, emphasising the need to stop or finalise within 72 h, and nine (23%) sites did neither.
Antibiotic use in the

Co-primary outcome: 30-day mortality
Preimplementation follow-up, months 33 (27-39; 20-41)   Postimplementation follow-up, months 22 (19-30; 16-38)   Co-primary outcome: total DDDs per general medical admission* Pre-implementation follow-up, months 33 (27-37; 20-41)   Post-implementation follow-up, months 23 (18-28; 14-37)   Data are median (IQR; range) or n (%), unless otherwise specified.DDD=defined daily dose.*Site 3 was not included in analyses of antibiotic use (appendix p 29). †Includes two sites that shared hospital-level DDDs due to limitations posed by local pharmacy information systems (sites 22 and 30).Further details are given in the appendix (p 50).‡Antibiotics in this category can be considered either access or watch depending on indication.Since indication was unknown, they were analysed separately.§These data were not required for pilot sites, so this criterion was treated as achieved in the analysis for those sites.2F).
There was no evidence that immediate effects on total DDDs per admission at implementation (-0•5%, 95% CI -2•7 to 1•7, per additional fidelity criteria achieved) and on year-on-year trends after versus before intervention (-1•4%, -4•9 to 2•3, per additional fidelity criteria achieved) were associated with overall implementation fidelity (figure 4).Immediate reduc tions in total DDDs per admission were greater among sites with processes for ongoing audit and feedback in place by implementation (by -16•6%, 95% CI -28•5 to -2•8, relative to sites that did not have processes in place by implementation) and greater among sites that submitted postim plementation audit data within 4 weeks following imple mentation (by -8•3%, -15•1 to -1•0, relative to sites that did not submit audit data within 4 weeks).However, the relative reduction in immediate implementation effect among sites that submitted postimplementation audit data within 4 weeks was not observed after we adjusted for whether the site had a process in place for ongoing audit and feedback by the implementation date in a multivariate model (appendix pp 16-20).We found nonsignificantly greater sustained reductions in total DDDs per admission among sites that introduced ARK categories into the prescribing process by implementation than among sites that had not introduced ARK categories by implementation (by -11•5%, 95% CI -22•9 to 1•7, relative to the reference group; appendix p 16) and among sites with higher uptake of the online learning by implementation (ie, with ≥20 people per 100 acute beds completing the training) than among sites with lower uptake (by -9•9%, 95% CI -19•7 to 1•1, versus sited training <20 people per 100 acute beds).Mediumsized sites also had non-significantly greater reductions in DDDs at implemen tation (by -7•4%, 95% CI -14•6 to 0•5, relative to small sites), with evidence for sustained year-on-year increases (by 14•6%, 0•1 to 31•3, relative to small sites; appendix pp 16-20).
Sites contributed a median 24 months (IQR 19-30, range 16-38) of data for all-cause 30-day mortality (in or out of hospital) after implementation (appendix p 46). *One site implemented the intervention in the feasibility phase of the research.These data were published previously and were not included in the main analysis. 20These sites declined to participate after full review of study materials typically due to the resource implications of participation in the research, conflicts with local antimicrobial stewardship initiatives, or being unable to provide mandatory electronic health record data.‡The decision not to fund data collection in sites withdrawing before implementation was made because, as mortality was a non-inferiority comparison, it was more important to replace these sites than use resources collecting data from sites that never implemented the intervention and hence would show no intervention effect on mortality.§One pilot site introduced an electronic prescribing system 6 months after implementation leading to an immediate decline (>90%) in reported antibiotic defined daily doses, which prevented valid assessment of trends in antibiotic use after implementation.||Individual-level data were not analysed due to scarcity.Sites are identified numerically by the order in which they were randomised to implement.The targets for 70% of essential people and more than 20 staff per 100 acute beds to complete ARK training are arbitrary but were prespecified for the funder as part of trial agreements and as part of prespecified fidelity criteria.Audit data were unavailable at two of 39 hospitals (ie, sites 31 and 38) and these sites are excluded from panels D-F.Six hospitals (sites 13, 18, 21, 27, 28, and 30) were missing baseline audit data and are therefore excluded from panels E-F.

Figure 3: Effects of the ARK intervention
Immediate effect at implementation (A) and effect on sustained year-on-year trend after versus before implementation (B).The top part of each panel shows the antibiotic primary (bold) and secondary outcomes, and the bottom part shows the clinical primary (bold) and secondary outcomes.Effects that were adjusted only for the effects of COVID-19 are shown in green, and fully adjusted effects are shown in red when there was evidence of an association or otherwise in blue.ARK=antibiotic review kit.C difficile=Clostridioides difficile.DDD=defined daily dose.ICU=intensive care unit.IRR=incidence rate ratio (negative binomial regression).OR=odds ratio (logistic regression).SHR=subhazard ratio (competing risks regression).*Antibiotics are measured as DDDs per admission unless indicated otherwise.†Access or Watch depending on indication (analysed as a mutually exclusive category, since indication was unknown).‡Modelled without a COVID-19 adjustment (effects plotted in green are therefore unadjusted).

Figure 4: Effect on total antibiotic DDDs per admission
Overall immediate effect at implementation (A), overall effect on year-on-year trend after versus before implementation (B), immediate effect at implementation by implementation fidelity (C), and effect on year-on-year trend after versus before implementation by implementation fidelity (D).Sites are identified numerically by the order in which they were randomised to implement and are ordered by the number of fidelity criteria achieved (appendix pp 13-15).The size of the symbols in panels C and D reflects the precision of each estimate (inverse of the within-hospital variance).
Weights are from random effects analysis.IRR=incidence rate ratio.

Figure 5: Adjusted 30-day mortality
Overall immediate effect at implementation (A), overall effect on year-on-year trend after versus before implementation (B), immediate effect at implementation by implementation fidelity (C), and effect on year-on-year trend after versus before implementation by implementation fidelity (D).Sites are identified numerically by the order in which they were randomised to implement and are ordered by the number of fidelity criteria achieved (appendix pp 13-15).Weights are from random effects analysis.The size of the symbols in panels C and D reflects the precision of each estimate (inverse of the within-hospital variance).OR=odds ratio.
After vs before implementation OR We identified weak evidence for sustained reductions in 30-day mortality among sites that introduced the ARK categories into the prescribing process by implementation than among sites that did not (-7•2%, 95% CI -14•6 to 0•8), among sites implementing a hard stop than among sites implementing no soft or hard stop (-8•2%, -15•0 to -0•9), and among sites implementing in July-September than among sites implementing in January-March (-10•3%, -19•0 to -0•8; appendix pp 16-20).
We did not identify any evidence that sites with greater reductions in antibiotic DDDs per admission had larger immediate (r=0•044, p=0•79) or sustained (r=0•011, p=0•95) increases in 30-day mortality trends than did sites with smaller reductions in antibiotic DDDs per admission (figure 6).

Discussion
Here, we have evaluated the ARK intervention, 18,20 which aimed to safely reduce antibiotic consumption in adult acute general medical hospital admissions, in a steppedwedge cluster-randomised trial.In our final model adjusting for COVID-19, the ARK intervention resulted in mean reductions in antibiotic use of 4•8% per year, but no immediate reduction.That the intervention changed prescribing over time rather than suddenly might be expected, given the different com ponents, including training for use of the novel decision aid and audit and feedback to re-enforce learning. 24The change over time could also reflect increasing acceptance that completion of arbitrary antibiotic courses might not reduce risk of resistance. 25Although the trial was powered to detect a 15% immediate reduction associated with the intervention, the effect observed is potentially clinically significant given that the national standard contract for acute trusts in England sought a reduction of only 1% per year.Given the importance of sustainable effects from behaviour change interventions in antibiotic stewardship, it is notable that this reduction was seen over a median of 23 months (range 14-37).Notably, consistent reductions were seen in Access, Watch, narrow-spectrum, and oral antibiotics, but not in broad-spectrum or parenteral antibiotics, antibiotics considered Access or Watch depending on indication, and piperacillin-tazobactam, and there was a significant increase in DDDs for carbapenems and Reserve classes.Since the intervention was targeted at acute general medical admissions, unsurprisingly its effect was seen in narrow-spectrum and Access agents, which are typically used as first-line medication or for de-escalation.The significant increase in carbapenem use after intervention could suggest that decreased use of one set of agents increased use of others.This effect seems unlikely, because the differences measured are relative and the absolute increases are small (appendix pp 49-50).Furthermore, use of carbapenems increased disproportionately across the NHS during the study period, driven by their inclusion in national treatment guidelines for hospital-acquired pneumonia, shortages of piperacillin-tazobactam, and increasing resistance to other agents. 15Broad-spectrum agents, such as carbapen ems, are typically prescribed when other agents have already been tried for the patient or when microbiology has identified a specific pathogen, and we might simply have observed an increase that the intervention would not be expected to affect.
We found no overall relationship between fidelity of implementation and the effect of the intervention.An absence of relationship might be because complex interactions between intervention elements and the implementation setting are difficult to measure quanti tatively in a large-scale trial, or because we took an average of how many fidelity criteria were met, but some of the criteria were likely to have had more of an effect on fidelity than others.The ARK audit tools were designed to support frequent, light-touch feedback to prescribers, sometimes called hand shake stewardship, 26 which relies on inter personal factors that we could not analyse but will be considered in forthcoming mixed-methods process analyses.Prescrip tion audits began 12 weeks before intervention to generate baseline data for the intervention's feedback element, so it is perhaps not surprising that rates of audit completion were generally higher before implementation than after wards.Notably, among individual intervention components, implementing the decision aid into the prescribing process and greater uptake of the online learning were both associated with greater reductions in antibiotic use than were not implementing these elements of the intervention, suggesting that these are key elements in achieving sustained change.
The ARK intervention focuses on decisions to stop rather than decisions to start antibiotics, because this approach has the potential to reduce overall use without withholding empirical antibiotics from patients with acute illness.Nevertheless, we considered it important to evaluate whether introducing ARK was associated with excess mortality.Beginning in March, 2020, when 12 of 39 sites were still within 12 months of implementation, the COVID-19 pandemic was associated with substantial increases in mortality among acute hospital admissions (appendix pp 51-127).Adjusting for this effect, in most of the main models and through sensitivity analysis excluding these 12 sites, we identified no clear evidence of associations between the intervention and 30-day or 90-day mortality.Notably, implementing the decision aid with a hard stop of antibiotic prescriptions at 72 h if not revised was associated with decreased risk of death over time, despite prescribers reporting anxiety that hard stops could compromise clinical outcomes. 27This decrease might be explained by clinicians placing a greater emphasis on prescription reviews at sites that introduced hard stops, improving patient management more broadly.Further more, we found no evidence that sites that achieved greater reductions in antibiotic DDDs per admission had larger increases in mortality than did sites with smaller reductions in antibiotic DDDs per admission (figure 6).
Our study has important limitations.First, there are intrinsic limitations of the cluster-randomised design.Although we included over a quarter of all acute hospitals in the UK health system, we cannot reliably exclude imbalance, particularly of time-dependent factors, as emphasised by the onset of the COVID-19 pandemic during the postimplementation period.There could be imbalance in other time-dependent organisational changes (eg, staffing, clinical or stewardship practice, or case-mix), which might have changed antibiotic consumption at individual sites.We do not have data for antibiotic resistance rates, which might have varied between sites over time and are generally lower in the UK than in many other countries.
Second, although sites were robustly randomised with respect to the timing of intervention implementation, they might not be a random sample of UK acute hospitals.It is plausible that only sites with well constituted antimicrobial stewardship teams volunteered, and other sites might not see the same effect, particularly as effect was associated with some aspects of intervention fidelity.Alternatively, the intervention effect could be greater at sites with weaker stewardship teams.
Third, we measured antibiotic consumption indirectly from dispensing data to clinical areas, as individual-level antibiotic data could be provided by only four sites.This method means that we cannot explore mechanisms through which the intervention reduced antibiotic use.However, our mortality analysis included over 7 million admissions, so there was no ability to collect individual prescribing data other than electronically.Although richer data at the individual-patient level would have allowed more detailed exploration, collecting consent and antibiotic use data from the number of patients needed to conduct a robust analysis would be infeasible.Furthermore, stewardship interventions, such as ARK, are made at the organisation level and, as such, organisation-level antibi otic use is an appropriate outcome.
Fourth, it is probable that not all prescribing decisions in the patient population analysed were subject to the intervention (eg, outlying surgical patients).Conversely, some patients for whom prescribing decisions were not subject to the intervention might have been included in analysis.These inclusions and exclusions are because acute general medical inpatients are not easily iden tified in electronic admission data, and we had to infer this population from specialty codes, which are used slightly differently across sites.Importantly, both these effects, and low implementation fidelity at some sites, would be expected to dilute the observed effect of the intervention on antibi otic use, suggesting that antibiotic reductions might have been even greater in targeted patients and in sites with high implementation fidelity.
In terms of potential clinical harms from the intervention, analysing routinely available electronic health records, we identified no consistent evidence of effect on mortality, admission to critical care, length of stay, or readmission.Although we cannot exclude the possibility of other harms related to shorter antibiotic treatment, our overall findings make substantial increases in treatment failure and recurrence unlikely.Equally, we were not able to measure potential direct benefits from reduced antibiotic treatment, but it is a reasonable assumption that reductions in antibiotic exposure will reduce antibiotic-associated harms, including resistance.
Despite its limitations, the cluster-randomised approach that we adopted allowed us to capture both the organ isation-level effects of the intervention on antibiotic consumption and the patient-level effects on clinical outcomes.Our findings are entirely consistent with the three, much smaller, previous trials of hospital stewardship interventions, which showed the importance of intervention co-design with practitioners, 28 practitioner education, and clinically relevant audit and feedback to clinicians. 29,30Our findings are also consistent with conclusion of the most recent Cochrane review that stewardship interventions can reduce unnecessary antibiotic use safely. 11Our approach to intervention design and evaluation addresses many of the limitations that have prevented the translation of previous research findings into hospital practice. 7,8Crucially, the wider ARK-Hospital programme has delivered practice-ready materials for implementation, which are freely available.Acute hospital providers should consider embed ding the ARK-Hospital toolkit in their staff training, pre scribing processes, and stewardship work to reduce antibiotic overuse in acute general medical inpatients and protect these patients from antibiotic-related harms.

Data sharing
The de-identified patient-level electronic health records (on over 7 million admissions) and hospital-level antibiotic use data used for this analysis was obtained from individual hospital organisations without permission for onward data sharing.It can be accessed either directly from the participating organisations or through the trial team if the participating organisations provide permission.De-identified patientlevel admission data can also be accessed directly through an application to NHS Digital.All enquiries should be sent to Prof Martin J Llewelyn (m.j.llewelyn@bsms.ac.uk or Prof Sarah Ann Walker  (sarah.walker@ndm.ox.ac.uk).The full protocol is available on http://www.arkstudy.ox.ac.uk/ark-for-healthcare-professionals.The statistical analysis plan is available by emailing Prof Martin J Llewelyn or Prof Sarah Ann Walker.

Figure 1 :
Figure 1: Flow diagram of participating hospital organisations*One site implemented the intervention in the feasibility phase of the research.These data were published previously and were not included in the main analysis.20†These sites declined to participate after full review of study materials typically due to the resource implications of participation in the research, conflicts with local antimicrobial stewardship initiatives, or being unable to provide mandatory electronic health record data.‡The decision not to fund data collection in sites withdrawing before implementation was made because, as mortality was a non-inferiority comparison, it was more important to replace these sites than use resources collecting data from sites that never implemented the intervention and hence would show no intervention effect on mortality.§One pilot site introduced an electronic prescribing system 6 months after implementation leading to an immediate decline (>90%) in reported antibiotic defined daily doses, which prevented valid assessment of trends in antibiotic use after implementation.||Individual-level data were not analysed due to scarcity.

Figure 2 :
Figure 2: Intervention adherence during the first 12 weeks of implementation Panels show the proportion of essential people who completed ARK training (A), staff who completed ARK training (B), staff who completed training per 100 acute beds (C), proportion of antibiotic prescriptions categorised using the decision aid at the initial prescription (D), proportion of antibiotic prescriptions reviewed versus baseline (E), and proportion of antibiotic prescriptions stopped at review and revise versus baseline (F).Sites are identified numerically by the order in which they were randomised to implement.The targets for 70% of essential people and more than 20 staff per 100 acute beds to complete ARK training are arbitrary but were prespecified for the funder as part of trial agreements and as part of prespecified fidelity criteria.Audit data were unavailable at two of 39 hospitals (ie, sites 31 and 38) and these sites are excluded from panels D-F.Six hospitals (sites13, 18, 21, 27, 28, and 30) were missing baseline audit data and are therefore excluded from panels E-F.
on-year change in trend before vs after implementation

Figure 6 :
Figure 6: Comparison of intervention effects on 30-day mortality and total antibiotic DDDs per admissionImmediate effect at implementation (A) and effect on year-on-year trend after versus before implementation (B).DDD=defined daily dose.IRR=incidence rate ratio.OR=odds ratio.

Contributors
TEAP, ASW, LY, and MJL conceived the research.ELAC, KS, SW, MSa, AK, FM, KSH, DWC, LV, SH LY, TEAP, and ASW conceived and developed the intervention.ML-S, RA, SB, PC, GC-B, SD, ME, RF, KJF, VG-A, SG, CG, KG, CH, DH, TH, SI, AJ, NJ, PK, GK, DMac, CM, DMaw, BM, MM, RM, SN, AN , JN, JO, AP, RP, NP, DP, ES, MSl, BS, CW, IW, MD, and MJL conducted the trial.EPB conducted the statistical analysis.EPB and ASW accessed and verified all the data.ASW, EPB, and MJL wrote the first draft.All authors reviewed and approved the final manuscript.All authors had full access to the data in the study and had final responsibility for the decision to submit for publication.Declaration of interestsMJL, DWC, LY, TEAP, and ASW declare funding from the National Institute for Health Research (NIHR) for the ARK-Hospital programme.ASW is an NIHR Senior Investigator.All other authors declare no competing interests.