Supplementing the National Early Warning Score (NEWS2) for anticipating early deterioration among patients with COVID-19 infection

Importance: An early minimally symptomatic phase is often followed by deterioration in patients with COVID-19 infection. This study shows that the addition of age and a minimal set of common blood tests taken in patients on admission to hospital significantly improves the National Early Warning Score (NEWS2) for risk-stratification of severe COVID disease. Objective: To supplement the NEWS2 score with a small number of easily obtained additional demographic, physiological and blood variables indicative of severity of COVID-19 infection. Design: Retrospective observational cohort with internal and temporal held-out external validation. Setting: Acute secondary care. Participants: 708 patients admitted to an acute multi-site UK NHS hospital with confirmed COVID-19 disease from 1st March to 5th April 2020. Intervention: Not applicable. Main outcome and measures: The primary outcome was patient status at 14 days after symptom onset categorised as severe disease (WHO-COVID-19 Outcomes Scales 6-8: i.e. transferred to intensive care unit or death). 218 of the 708 patients reached the primary end point. A range of physiological and blood biomarkers were assessed for their association with the primary outcome. Adjustments included age, gender, ethnicity and comorbidities (hypertension, diabetes, heart, respiratory and kidney diseases). Results: NEWS2 total score was a weak predictor for severity of COVID-19 infection at 14 days (internally validated AUC = 0.628). The addition of age and common blood tests (CRP, neutrophil count, estimated GFR and albumin) provided substantial improvements to a risk stratification model but performance was still only moderate (AUC = 0.75). Common comorbidities hypertension, diabetes, heart, respiratory and kidney diseases have minor additional predictive value. Conclusions and relevance: Adding age and a minimal set of common blood parameters to NEWS2 improves the risk stratification of patients likely to develop severe COVID-19 outcomes. The addition of a few common parameters is likely to be much easier to implement in a short time-scale than a novel risk-scoring system.


Introduction
While approximately 80% of individuals with COVID-19 infection have mild or no symptoms 1 , some develop severe COVID-19 disease requiring hospital admission. As of 23rd April 2020, there have been >2.5 million confirmed cases worldwide 2 . Within the subset of those requiring hospitalisation, early identification of those who deteriorate and require transfer to an intensive care unit (ICU) for organ support or may die is invaluable 12 .
Currently available risk scores for deterioration of acutely ill patients include (1) widely-used generic ward-based risk indices such as the National Early Warning Score (NEWS2) 3 or modified sequential organ failure assessment (mSOFA) 4 ; and (2) the pneumonia-specific risk index, CURB-65 5 which usefully capture a combination of physiological observations with limited blood markers and comorbidities. The NEWS2 is a summary score of six physiological parameters or 'vital signs' (respiratory rate, oxygen saturation, systolic blood pressure, heart rate, level of consciousness, temperature and supplemental oxygen dependency), used to identify patients at risk of early clinical deterioration in the UK NHS hospitals 6,7 . The physiological parameters assessed in the NEWS2 score -particularly patient temperature, oxygen saturations and the supplemental oxygen dependency -have been associated with COVID-19 outcomes 1 ; however, little is known about their predictive value for the severity of COVID-19 disease. Additionally, a number of COVID-19-specific risk indices are being developed [8][9][10] as well as unvalidated online calculators 11 but generalisability is not yet known 10 . A Chinese study has suggested a modified version of NEWS2 with addition of age only 12 but without any data on performance. With near universal usage of NEWS2 in UK NHS Trusts since March 2019 13 , minor adaptation to NEWS2 would be relatively easy to implement.
As the SARS-Cov2 pandemic has progressed, evidence has emerged regarding potentially useful blood biomarkers 1,[14][15][16][17] . Although most of these early reports contain data from small numbers of patients, a number of markers have been found to be associated with severity. These include neutrophilia and lymphopenia, particularly in older adults 9,16,18,19 , neutrophil-to-lymphocyte ratio 20 , raised C-Reactive Protein (CRP) and lymphocyte-to-CRP ratio 20 , markers of liver and cardiac injury such as alanine aminotransferase (ALT), aspartate aminotransferase (AST) and cardiac troponin 21 and elevated D-dimers, ferritin and fibrinogen 2,5,7 . Furthermore, plasma levels of cytokines such as IL-6 have been found to be higher in COVID-19 patients compared to controls 1 .
Our aim is to understand the performance of NEWS2 and identify a supplemental combination of simple clinical and blood biomarkers routinely measured in hospitals to supplement the NEWS2 score to improve prediction of a severe disease outcome at 14 days from symptom onset. To reach this aim, our specific objectives were: 1. To explore independent associations of routinely measured physiological and blood parameters (including NEWS2 parameters) at or near hospital admission with disease severity (i.e., ICU admission or death), adjusting for socio-demographics and comorbidities. 2. To examine which minimal combination of these potential determinants of disease severity (physiological and blood parameters, sociodemographics and comorbidities) are the best predictors of disease severity at 14 days since symptom onset; and 3. To compare the predictive value of the resulting model with a model based on the NEWS2 total score alone.

Patients
The study cohort was defined as all adult inpatients testing positive for SARS-Cov2 by reverse transcription polymerase chain reaction (RT-PCR) between 1 st March to 5 st April 2020 at a multi-site acute NHS hospital in South East London (UK Blood parameters . We focused on biomarkers that were routinely obtained at or shortly after admission and were therefore available for the vast majority of patients. These comprised: albumin (g/L), alanine aminotransferase (ALT; IU/L), creatinine (µmol/L), C-reactive protein (CRP; mg/L), estimated Glomerular Filtration Rate (eGFR; mL/min), Haemoglobin (g/L), lymphocyte count (x 10 9 /L), neutrophil count (x 10 9 /L), and platelet count (PLT; x 10 9 /L). We also derived the neutrophil-to-lymphocyte ratio (NLR) and the lymphocyte-to-CRP ratio 13 . Troponin-T (ng/L) and Ferritin (ug/L) were included, although these measures were only available for a subset of participants. D-dimers and HbA1c were excluded since they were measured in very few patients at admission and insufficient samples were available for analysis.
Physiological parameters. We included the six physiological parameters that form the basis of the NEWS2 score, namely, respiratory rate (breaths per minute), oxygen saturation (%), systolic blood pressure (mmHg), heart rate (beats/min), temperature (°C), and consciousness (measured by Glasgow Coma Scale (GCS) total score). All were measured at or shortly after admission. We assessed these parameters individually as well as a NEWS2 total score. Diastolic blood pressure, which is not part of the NEWS2 score, was also included in the analyses.
Demographics and comorbidities. Age, sex, ethnicity and comorbidities were considered. Where ethnicity data was available this was categorised as caucasian vs. BAME (Black, Asian and minority ethnic). For supplementary models adjusting for ethnicity, patients with ethnicity reported as 'unknown/mixed/other' were excluded. We included binary measures (present vs. not present) of relevant comorbid chronic health conditions derived from the NLP pipeline described above: hypertension, diabetes, heart disease (heart failure and ischemic heart disease), respiratory disease (asthma and chronic obstructive pulmonary disease, COPD) and chronic kidney disease .

5
. CC-BY-NC-ND 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) The copyright holder for this preprint this version posted April 29, 2020. . https://doi.org/10.1101/2020.04. 24.20078006 doi: medRxiv preprint

Statistical analyses
Preliminary descriptive and exploratory analyses were performed. To address our first objective -exploring independent associations of physiological and blood parameters with 14-day death/ICU -we used penalised maximum likelihood logistic regression which reduces bias due to small sample size 27 . Each parameter was tested independently, adjusted for age and sex (Model 1) and then additionally adjusted for comorbidities (Model 2). Parameters exhibiting skewed distributions were transformed before modelling with logarithmic or square-root transformations. All parameters were scaled (mean = 0, standard deviation = 1) to improve interpretability. Outlying high values for some blood parameters were retained after individual examination by clinicians who ascertained their plausibility. We used the maximal available sample when testing each parameter. Given the number of tests conducted, P -values were adjusted using the Benjamini-Hochberg procedure to keep the False discovery rate at 5% 28 . These models were conducted with R 3.6 23 using the logistf 24 package.
To address our second and third objectives -which combination of parameters performed best in predicting the 14-day outcome over and above NEWS2 -we estimated models combining all parameters using regularized logistic regression with a LASSO (Least Absolute Shrinkage and Selection Operator) estimator which shrinks parameters according to their variance, reduces overfitting and enables automatic variable selection 29 . The optimal degree of regularization was determined by identifying a tuning parameter λ using cross-validation 30 . LASSO regression provides a sparse, interpretable model, which allows us to predict individual risk scores (i.e. probability of severe outcome). Starting from an initial model with NEWS2 total score only, sets of features were added in order of (i) age and sex, (ii) blood and physiological parameters; (iii) comorbid conditions. A final model was estimated using NEWS2 total score alongside the top five most influential features from previous models. To estimate the predictive performance of our model on new unseen cases of the same underlying population, we performed internal nested cross-validation (10 folds and 20 repeats for the inner loop; 10 folds and 100 repeats for the outer loop). Overall discrimination was assessed based on the area under the curve (AUC). All continuous features were scaled (mean = 0, standard deviation = 1). Missing feature information was imputed (after scaling) using k-Nearest Neighbours imputation (k=5).Scaling and kNN imputation were incorporated within the model development and selection process to avoid data leakage which would otherwise result in optimistic performance measures 31 .
To assess whether a more complex machine learning estimator would improve predictive performance, we repeated this set of models using gradient boosted trees implemented in the XGBoost library 32 . Procedures for internally validating these models were equivalent to those described above for regularized logistic regression except the imputation step was omitted due to the ability of XGBoost to handle missing data.
The predictive performance of the derived regularized logistic regression model was then evaluated by temporal external validation 33 with a hold-out sample of 256 patients who were admitted to hospital after the training sample (see Supplementary Figure 1). This involved 6 . CC-BY-NC-ND 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)
The copyright holder for this preprint this version posted April 29, 2020. . estimating the original model exactly as presented, including scaling and imputation models derived in the training data set. Discrimination performance was assessed using AUC, sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV). Model calibration was assessed using a calibration plot (model predicted probability vs. true probability). These models were estimated in Python 3.6 34 using NumPy 35 , and Scikit-Learn 36 .
Sensitivity analyses were performed to account for potential demographic variability. Recent evidence suggest sex differences with men more likely to experience worse outcomes 16 . Therefore, in separate models, we tested interactions between each physiological and blood parameter and sex using likelihood-ratio tests (comparing a null model with the main effects only vs. a model additionally including the interaction term). In addition, we replicated all models with adjustment for ethnicity in the subset of individuals with available data for ethnicity (n=285 in training sample).

Results
The initial inpatient cohort comprised 452 inpatients testing positive for COVID-19 of whom 159 (35%) were transferred to ICU or died (COVID-19 WHO Score 6-8) within 14 days of symptom onset. Table 1 describes the clinical characteristics of the cohort: the mean age was 67 years (standard deviation = 18.5); 54% (n=248) were male; 42% (n=120) were categorised as BAME. Patients associated with a more severe outcome were significantly older (71 vs. 65 years; p = 0.004) but there was no evidence of differences by sex or ethnicity. There were some differences between groups in the prevalence of comorbidities but these did not reach statistical significance after multiple testing correction. For example, compared to patients with less severe outcomes, those who transferred to ICU or died had higher rates of hypertension (60% vs. 50%; p = 0.11), diabetes (38% vs. 32%; p = 0.33), heart failure (16% vs. 11%; p = 0.33) and chronic kidney disease (24% vs. 16%; p = 0.11). Rates of other comorbidities were similar between the two groups. There were differences between outcome groups for most blood and physiological parameters. Patients who had transferred to ICU or died within 14 days had, at admission, lower levels of Albumin, ALT, and estimated GFR; and elevated levels of CRP, creatinine, Ferritin, and Neutrophils. Mean NEWS2 total scores were significantly different (3.4 vs 2.1; p < 0.001; corresponding to Cohen's d of -0.57) in patients who transferred to ICU or died, compared to inpatients experiencing less severe outcomes.
Logistic regression models were used to assess independent associations between each physiological and blood parameter and disease severity measured as transfer to ICU or death ( Table 2). Individuals were more likely to have transferred to ICU/died within 14 days of symptom onset if: they had higher CRP, NEWS2 score, heart rate, neutrophils, neutrophil-lymphocyte ratio, respiration rate; or if they had lower lymphocyte/CRP ratios, eGFR, creatinine, and oxygen saturation. These associations remained after adjustment for age, sex and comorbidities. There was no evidence of differences by sex (results not presented) and findings were consistent when additionally adjusting for ethnicity in secondary analyses using the subset of individuals with ethnicity data (Supplementary Table 3).

7
. CC-BY-NC-ND 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)
The copyright holder for this preprint this version posted April 29, 2020. . https://doi.org/10.1101/2020.04. 24.20078006 doi: medRxiv preprint Combining physiological and blood parameters to assess ability to improve on NEWS2 in predicting 14-day outcome To identify which minimal set of parameters were best able to improve on NEWS2 in predicting the 14-day outcome (ICU/death vs. not ICU/death), we combined all predictors in a single logistic regression model using LASSO regularisation. Internally validated predictive performance based on the area under the ROC curve (AUC) is presented in Table 3 for different feature sets. NEWS2 shows poor discrimination with an AUC of 0.628. Adding age and sex to a baseline model of NEWS2 total score only increased the AUC by 0.025 to 0.653 (+/-2SD range: 0.639, 0.667). Further adding in all other blood and physiological parameters (except NEWS2) increased the AUC further by 0.089, to 0.742 (+/-2SD: 0.726, 0.758). Additionally including comorbidities in this model did not improve performance. A final model was estimated including NEWS2 and the top five most important features taken from Model 4. This simpler model resulted in a slightly larger AUC of 0.751 (+/-2SD range: 0.737, 0.764) which may indicate some overfitting due to the pre-selection of variables from previous analyses. Results were consistent when repeating these models in the subset of patients with information available on ethnicity (Supplementary Table 5). Figure 1 summarises feature importances from the LASSO logistic regression models. When adding blood and physiological parameters to NEWS2 ('NEWS2 + DBP'), 8 features were retained, in order of effect sizes: NEWS2 total score, CRP, neutrophils, estimated GFR, albumin, age, Troponin T, and oxygen saturation. Notably, when additionally considering comorbid conditions ('NEWS2 + DBPC'), the retained features were similar, and no comorbid conditions were retained. This suggests that most of the variance is already captured by the top 5 parameters.
When these models were repeated using a more complex estimator (gradient boosted trees, using XGBoost 32 ) the pattern of results was consistent with those from regularized logistic regression (Supplementary Table 5). Namely, the internally validated AUC improved from 0.646 for a model with NEWS2 alone, to 0.722 for a model that additionally included the five parameters: CRP, neutrophils, estimated GFR, albumin, and age. Importantly, while the pattern of results was consistent, a more complex machine learning estimator produced no improvements to predictive performance.
Temporal external validation was conducted on a hold-out sample of 256 patients. This sample was similar to the training sample on all parameters (Supplementary Table 6) except the proportion who transferred to ICU or died was lower. Overall, results from the hold-out sample were consistent with those from internal validation. The AUC for NEWS2 alone was 0.700, and this improved to 0.730 when adding all blood and physiological parameters (sensitivity = 0.441; specificity = 0.873). The AUC for the simplified final model including NEWS2 and the top five features (CRP, neutrophils, estimated GFR, albumin and age) was similar (AUC = 0.730; sensitivity = 0.458; specificity = 0.873) (Supplementary Table 7). Calibration for these models 8 . CC-BY-NC-ND 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.

Discussion
To our knowledge our study is the first to systematically attempt to improve performance of NEWS2 specifically for COVID-19. We found that the NEWS2 score shows overall poor discrimination with high specificity but poor sensitivity for severe outcomes in COVID-19 infection (transfer to ICU or death). However, its value for risk stratification (especially sensitivity) can be significantly improved by adding age and a small number of additional blood parameters (CRP, neutrophils, estimated GFR and albumin). A number of blood measures previously linked with more severe outcomes -such as lymphocyte and ALT 14 , or transformations of inflammatory markers such as CRP/lymphocyte or neutrophil/lymphocyte ratio -did not provide additional value to the model over and above the existing features despite being more common in those individuals with more severe outcomes. Moreover, cardiac disease and myocardial injury has been described to be commonly seen in the severe COVID-19 cases in China 1,21 . In our model, blood Troponin-T, a marker of myocardial injury, had additional salient signal but was only measured in a subset of our cohort at admission, so it was not included in our final model. This would have to be explored further in larger datasets. A systematic review of 10 prediction models for mortality in COVID-19 infection 10 found broad similarities with the features retained in our models, particularly regarding CRP and neutrophil levels. However, existing prediction models suffer several methodological weaknesses including over-fitting, selection bias, and reliance on cross-sectional data without accounting for censoring. Additionally, almost all existing studies have relied on ethnically homogenous Chinese cohorts and thus may be unrepresentative of other global populations.
With regards to pre-existing disease comorbidities (hypertension, diabetes mellitus, heart failure, ischaemic heart disease, COPD, asthma and chronic kidney disease), these were more common in patients with severe outcomes but had minimal contribution to the risk prediction and were not retained in the final model. This was unexpected and suggests potential shared variance between pre-existing health conditions and some of the included blood or physiological markers. Future research should explore further the potential underlying shared mechanisms that can predict deterioration.
NEWS2 is a summary score derived from six physiological parameters, including oxygen saturation. While NEWS2 total score was one of the most influential parameters in our models, the oxygen saturation sub-parameter remained influential and was retained following regularisation (i.e. model 'NEWS + DBP'). This suggests some residual association over and above what is captured by the NEWS2 score between oxygen saturation and more severe outcomes, and reinforces Royal College of Physicians guidance that the NEWS2 score ceilings with respect to respiratory function 37 .

9
. CC-BY-NC-ND 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.

Strengths and limitations
Our study included data from a large sample of patients admitted to hospital with high rates of the primary outcome (transfer to ICU or death) and considered a large number of potential predictors including demographics, physiological and blood parameters and comorbidities. However, some limitations should be acknowledged. First, there are likely to be other parameters not measured in this study that could improve the risk stratification model substantially (e.g. radiological features, other comorbidities or comorbidity load). This could be addressed by future work to introduce additional data modalities, but these were not considered in the present study to avoid limiting the real-world implementation of the risk stratification model; a complex model with many parameters will be harder to implement in clinical practice. Second, we used a 14-day time window from the symptom onset date as this provides a balance between medium-term prognostication and actionable risk stratification at the usual period of deterioration. Longer timeframes may be useful for prognostication but are harder to generalise due to the greater number of factors affecting outcomes, including institutional, regional or national policies. Since NEWS2 score is optimised for very near-term deterioration at 24 hours 7 , a 14-day window was used as a compromise. Third, while the hold-out sample used for temporal external validation was similar in terms of demographics, blood and physiological parameters, the rate of more severe outcomes differed significantly. Perhaps due to changes in hospital procedures over time, this again suggests the need to validate these models in other hospitals or regions. Finally, while the model was derived from two hospital sites providing a mixed population, this study highlights that initial prediction models still have poor sensitivity and recalibration would be required before implementation as a risk model in clinical practice. Validation across datasets from a wider geographical region will be necessary to ensure generalisability.

Conclusion
In conclusion, this study suggests that the simple addition of a limited number of blood parameters to the existing and widely implemented NEWS2 system can contribute to improved risk stratification among COVID-19 patients. Our model can be easily implemented in clinical practice and predicted risk score probabilities of individual patients are easy to communicate. The additional parameters are widely collected on patients at hospital admission, and with near universal usage of NEWS2 in NHS Trusts since March 2019 13 , a minor adaptation to NEWS2 is substantially easier to implement in a variety of health settings than a bespoke risk score.

10
. CC-BY-NC-ND 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.

This paper represents independent research part funded by the National Institute for Health Research (NIHR) Biomedical Research Centres at South London and Maudsley NHS Foundation Trust, and Guy's & St Thomas' NHS Foundation Trust, both with King's College
London. The views expressed are those of the author(s) and not necessarily those of the NHS, the NIHR or the Department of Health and Social Care. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. We would also like to thank all the clinicians managing the patients, the patient experts of the KERRI committee, Professor Irene Higginson, Professor Alastair Baker, Professor Jules Wendon, Dan Persson and Damian Lewsley for their support.
12 . CC-BY-NC-ND 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)
The copyright holder for this preprint this version posted April 29, 2020. 15 . CC-BY-NC-ND 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)
The copyright holder for this preprint this version posted April 29, 2020. 16 . CC-BY-NC-ND 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)
The copyright holder for this preprint this version posted April 29, 2020

17
. CC-BY-NC-ND 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)
The copyright holder for this preprint this version posted April 29, 2020. . 18 . CC-BY-NC-ND 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)
The copyright holder for this preprint this version posted April 29, 2020 Notes. 1 FDR-corrected P-values based on the Benjamini-Hochberg correction.
Odds ratios represent a one standard deviation change in the respective blood and clinical measure at admission (tested in separate models). Model 1 adjusted for age and sex. Model 2 additionally adjusted for comorbidities (hypertension, diabetes, heart diseases, respiratory diseases and chronic kidney disease).

19
. CC-BY-NC-ND 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)
The copyright holder for this preprint this version posted April 29, 2020.

24
. CC-BY-NC-ND 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)
The copyright holder for this preprint this version posted April 29, 2020.

25
. CC-BY-NC-ND 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) The copyright holder for this preprint this version posted April 29, 2020.

26
. CC-BY-NC-ND 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) The copyright holder for this preprint this version posted April 29, 2020. . . CC-BY-NC-ND 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) The copyright holder for this preprint this version posted April 29, 2020. 28 . CC-BY-NC-ND 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) The copyright holder for this preprint this version posted April 29, 2020.

29
. CC-BY-NC-ND 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) The copyright holder for this preprint this version posted April 29, 2020. . https://doi.org/10.1101/2020.04. 24.20078006 doi: medRxiv preprint