Human infection due to the novel severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2)1 (commonly referred to as COVID-19 disease) has had unprecedented global impact on society and healthcare provision. In the first seven months of 2020, there have been more than 600,000 fatalities, with three countries (USA, Brazil and UK) each reporting more than 40,000 deaths.2

Current management protocols have generally emerged from China, where risk factors for in-hospital death included older age, higher Sequential Organ Failure Assessment (SOFA) score, and D-dimer greater than 750ng/mL on admission.3 Cytokine patterns3–5 were recognized as compatible with a “cytokine storm” of dysregulated inflammation and hyperpyrexia.6–9 Adult respiratory distress syndrome (ARDS), acute kidney, cardiac and liver failure, and a cytokine storm driving later deteriorations have been foci of research and clinical initiatives, alongside development of vaccines and antiviral therapeutics.10 Autopsy studies11–14 provided a slightly different picture as the risk of fatal thromboembolic disease emerged, in keeping with the initial reports of elevated D-dimer associations.3 Thromboembolic disease is now estimated to affect high proportions of patients with COVID-19.15–18

COVID-19 displays heterogenous presentation patterns and outcomes. Patient comorbidities including hypertension, diabetes, obesity, and heart disease are associated with higher risk of severe COVID-19 disease.3,4,19–21 while patients with additional common diseases including asthma and chronic obstructive lung disease (COPD) are advised to take extra precautions in the USA.22 Reported COVID-19 inpatient mortality rates vary by city/region,23 between countries,1,3,4,16,17,19,23–25 and by disease severity.3,4,23–26 Differential survival rates can therefore reflect a multitude of confounders from patients, geography, and health care systems, superimposed on pathophysiological variability that is not yet understood.

A near real-time decision tool was introduced early in the pandemic in our institution, as we were concerned about the high published mortality rates. This tool refreshes in the background every 10min and updates a clinical dashboard accordingly. Generated on the electronic patient record, traffic lights were used to alert clinicians when patients met criteria for increased risk of the three most common causes of death in COVID-19 – thromboembolism, cytokine storm, and respiratory deterioration/ARDS.

The goal of the current study was to evaluate whether prospective application during April and May 2020 had been associated with any change in patient early mortality across completed hospital episodes.

We have shown that early introduction of near real time traffic light-driven, personalized care was associated with an overall reduction in crude mortality from 32.3% (95% confidence interval 28.0, 36.0), to 23.0% (95% confidence interval 17.1, 28.9). The overall age and sex-adjusted probability of mortality reduced from 0.33 (95% confidence intervals 0.30, 0.37), to 0.22 (95% confidence interval 0.17, 0.27), p<0.001. This was despite the later cohort including more patients with thromboembolism flags associated with poorer survival. The greatest benefits were observed in older patients, males, and in patients with hypertension (p-values ≤0.01) and/or diabetes (p=0.05). Taken together, these data suggest that the early clinical interventions modified progression of COVID-19 induced disease, and reduced the proportion of patients progressing to critical illness and death.

There were several study limitations with many shared by other studies in emergency-based assessments of this new human disease with very high mortality. Interventions prioritized clinical care delivery, and the main purpose of this manuscript was to stratify the disease and introduce early medical management, rather than detailed expositions of population characteristics and symptomatology. Patients with very prolonged hospitalization did not have completed encounters and were not included: responses in this important subgroup will need evaluation in future prospective randomized controlled trials. Additionally, capturing mortality to end-of-hospital admission may have missed subsequent deaths in the community, and these may have differed in proportions over the study period. However, for crude mortality in the “Post” cohort to reach that of the “Pre” cohort, an 86% mortality rate would be required in the inpatients admitted on/after 12th April 2020, when their current length of stay IQR already exceeds the IQR for length of stay before death (Supplementary Table A.1). Therefore, this reduction in early mortality is considered to be robust, and important to communicate.

The study was conducted at a two-site institution, therefore it is difficult to compare baseline mortalities with other institutions. Lower baseline mortality was reported for 2569 completed episodes in a different health care setting in New York, where 665/2569 (25.9%) patients died, though highlighting the potential problems of comparing different outcomes in different institutions, the New York cohort24 were ∼6 years younger, and their interquartile range (IQR) for D-dimer did not overlap with the current cohort (New York IQR 237–714ng/mL, current cohort 818–2746ng/mL).

With the very high worldwide mortality2 and strong steers toward risk factors from the earliest reports from China,1,3 it was not considered appropriate to randomize patients. Where lethality is high, as in COVID-19, the paucity of sequential analysis techniques has been noted with concern.28 The current study employed a sequential analysis technique, within a dual center, single institution, and demonstrated that the introduction of full traffic light-driven care on 12th April 2020 was associated with a 33% reduction in age and sex-adjusted inpatient mortality. We recognize that a limitation of a sequential analysis is that it is not possible to fully standardize variables, and acknowledge that other factors could have contributed. It is feasible that in the current study, the baseline “Pre” group were already benefiting from the first phase of traffic light implementation,27 and different approaches to care.29 The survival signal was robust to age and sex adjustment; adjustment for patient severity indices would have made the association stronger, and other major treatment protocols did not change in the study period. Clinical expertise and pathways did have extra time to embed, though we note unchanged mortality figures in major country-wide audits over the same period.30 It is therefore difficult to conclude that any other factor could have been responsible for generating a completed encounter inpatient mortality rate that at 23.0% [95% confidence intervals 17.1, 28.9] is now lower than in New York,24 and lower than the rates reported during the Middle East Respiratory Syndrome (MERS) epidemic.31 Further, this completed encounter mortality rate was achieved with more than 25% of the patient population aged 81 years or over at the time of admission, and with advanced disease at presentation as indicated by the interquartile ranges for admission CRP (40–171mg/L), admission D-dimer (818–2746ng/mL) and admission ferritin (245–1100μg/L).

Despite more severe disease in the “Bridge” cohort, survival rate was no worse than for the “Pre” cohort. This suggests that the “Bridge” cohort may have also benefitted in the first phase of traffic-light implementation.27

Numerically, the key differences compared to our initial practice were earlier, enhanced anticoagulation for patients identified as at higher risk of thromboembolism disease. A smaller group of patients flagged at risk of cytokine storm based on serum ferritin also had enhanced attention to fluid balance and intravenous fluid delivery. For thrombosis, the flags helped to identify the stratified risk, and the intervention was based on the risk regardless of the radiological findings. In COVID-19, thrombosis is commonly microvascular and difficult to detect using current imaging, with clinical reports tending to focus on the presence or absence of major thrombus. The observed benefit from enhanced anticoagulation is in keeping with the limited published evidence in COVID-19.26,32 The prothrombotic COVID-19 state is currently considered less to reflect disseminated intravascular coagulation (DIC) caused by the systemic generation of thrombin, than hypercoagulability in the setting of a severe inflammatory state,33 and particularly in situ thrombus formation.34–36 The findings would also be supported by previous evidence on multi-organ failure in sepsis where inhibiting thrombin generation by anticoagulation may have benefit in reducing mortality.37

There was early evidence that the presence of an elevated D-dimer was associated with a poor prognosis in COVID-19.3,38 It has been suggested that high D-dimer values may not reliably predict the presence of a thrombus that has been degraded by fibrinolysis, but instead represent a marker of poor overall outcome in COVID-19.39 However, data from the current study targeting D-dimer-identified subgroups, and from two other groups,26,32 indicate that enhanced anticoagulation can improve survival, in advance of full mechanistic understanding.

At the time of revised manuscript submission, there remain no data on where the risk of thromboembolism starts to rise on the scale of elevated D-dimer. Interpretation of the stepwise accrual of survival benefit with heparin treatment as D-dimer exceeded the upper limit of normal26 varies, as demonstrated by the six different classifications for high D-dimer used as inclusion criteria for COVID-19 anticoagulation studies listed on clinicaltrials.gov. Combining elevated D-dimer with other biomarkers has been proposed for thromboembolic risk stratifications in COVID-19 patients.40 As proposed in,40 we also examined fibrinogen initially, and unexpectedly found that CRP (which was not included in many Chinese reports),3,4 was more discriminatory. CRP and D-dimer not only identified higher risk categories for enhanced treatments, but also a low risk subgroup (all D-dimer ≤1000ng/mL, and all CRP ≤200mg/L). The ability to identify subgroups with better prognosis is generally helpful in establishing where risks of intensive therapies may be more difficult to justify. Since therapeutic anticoagulation carries hemorrhagic concerns, particularly for diseases with abnormal vasculature as in COVID-19,11–14,18 we speculate that restricting enhanced thromboembolic prophylaxis to patients in subgroups with higher thrombotic risks may have contributed to the overall favorable survival figures in the traffic lights-managed cohort.

In conclusion, early recognition of thromboembolism risk based on CRP and or D-dimer, and impending cytokine storm predominantly based on serum ferritin, enabled hospital survival for completed encounters of 77.0% (95% confidence intervals, 71.1–82.8%) when more than 25% of admitted patients were at least 81 years old. Many COVID-19 at-risk groups benefitted including older patients, males, and patients with concurrent hypertension and diabetes. The blood measurements are already part of most hospitals’ COVID-19 datasets, and interventions for thromboembolic disease, and fluid management are widely available and inexpensive. Thus broad application seems feasible, and wider implementation is encouraged.

The code and guidance is freely available on request from Alex.McCarthy@chelwest.nhs.uk, and is to be published to GitHub in due course. Fully anonymised data will be available post peer review publication on reasonable request, in accordance with institutional protocols above. A data dictionary defining each field in the data set will also be made available to others.

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Support for clinical care was provided by Chelsea & Westminster Hospital NHS Foundation Trust, London, UK, and the National Health Service (NHS) in England.

The study was conceived and overseen by MPV. Literature searches were performed by MPV, CLS, AM, AG, SP, and MH. The traffic light bioinformatics system was designed and implemented by MPV, AM, AH, and AB. The clinical study design was by MPV. Haematology care pathway design was by MPV, AG and SP. Data analysis for real-time feedback was performed by MPV and AM. Real time analytic data application was performed by MPV, AG, PS, AB, SC, PD, PL, SP, TP, MH, RTK, IB, and GD. Data analysis was devised and performed by CLS. Data interpretation was performed by MPV and CLS. Figures were generated by CLS. The manuscript was written by CLS, and revised by MPV, AM, SP, PLS, SC, and MH. All authors reviewed and approved the final manuscript..

The authors declare no conflicts of interest.

The authors wish to thank the Chelsea & Westminster NHS Foundation Trust personnel and especially the Adult Intensive Care Unit nurses, for the delivery of personalised care to all patients admitted with COVID-19 infection. The authors would also wish to thank the CW Plus charity for invaluable support throughout the COVID-19 outbreak.