The low‐harm score for predicting mortality in patients diagnosed with COVID‐19: A multicentric validation study

Abstract Objective We sought to determine the accuracy of the LOW‐HARM score (Lymphopenia, Oxygen saturation, White blood cells, Hypertension, Age, Renal injury, and Myocardial injury) for predicting death from coronavirus disease 2019) COVID‐19. Methods We derived the score as a concatenated Fagan's nomogram for Bayes theorem using data from published cohorts of patients with COVID‐19. We validated the score on 400 consecutive COVID‐19 hospital admissions (200 deaths and 200 survivors) from 12 hospitals in Mexico. We determined the sensitivity, specificity, and predictive values of LOW‐HARM for predicting hospital death. Results LOW‐HARM scores and their distributions were significantly lower in patients who were discharged compared to those who died during their hospitalization 5 (SD: 14) versus 70 (SD: 28). The overall area under the curve for the LOW‐HARM score was 0.96, (95% confidence interval: 0.94–0.98). A cutoff > 65 points had a specificity of 97.5% and a positive predictive value of 96%. Conclusions The LOW‐HARM score measured at hospital admission is highly specific and clinically useful for predicting mortality in patients with COVID‐19.

68er7) on April 21, 2020. The original acronym for this score (HOT CALL) was changed to avoid confusion with the preexisting CALL score. INTRODUCTION

Background
Multiple prognostic factors for disease severity in patients diagnosed with coronavirus disease 2019 (COVID-19) have been identified. [1][2][3] In this regard, many prognostic scores have already been put forward to predict the risk of death and other outcomes (eg, CALL score, ABC GOALS, Neutrophil-Lymphocyte index, etc). [4][5][6] However, hospitals in developing countries often cannot measure some of the variables included in these scores (D-dimer, ferritin, computed tomography [CT] scans, etc). Moreover, implementation of many of these scores is hampered by the inclusion of subjective variables such as breathlessness, 5 data on preexisting comorbidities 7 (making it impossible to reassess prognosis according to the patients' clinical evolution) or rely on cutoff values that are infrequently met by patients with COVID-19 in realworld settings.

Importance
Developing countries have a lower number of critical-care beds 8 and specialists per 100,000 people. 9 Thus, estimating mortality is essential for optimal resource allocation. Prediction tools also have ethical applications and implications. Some triage systems repurpose scores to predict mortality in critical care patients, such as the SOFA (Sequential Organ Failure Assessment) score, as part of their decision framework. 10,11 However, there is compelling evidence highlighting the importance of generating and using disease-specific prediction tools or models in pandemic contexts. 12 Mathematical models for estimating new cases of COVID-19 in the post-pandemic period agree there will be >1 "wave" of infections, 13 and serological surveys for estimating the dynamics of a population's susceptibility, level of exposure, and immunity to the virus support these predictions. [14][15][16] Therefore, an effective prognostic tool is still relevant even if most countries are already flattening their daily curve of confirmed cases. 17 Furthermore, having context-specific predictive accuracy is essential for assisting the decisionmaking process in these extraordinary situations, for objectively tracking clinical status, and for providing realistic and accurate information to patients and their families about prognosis.

Goals of this investigation
This work evaluated the predictive performance of the novel LOW-HARM score (Lymphopenia, Oxygen saturation, White blood cells, Hypertension, Age, Renal injury, and Myocardial injury) for predicting mortality in patients diagnosed with COVID-19.

Study design and setting
This work was an observational analytic cohort study. The project and analysis strategy were preregistered at the Open Science Framework

Selection of participants
We collected and analyzed data from all patients with severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection confirmed by RT-PCR that were consecutively hospitalized at the alreadymentioned institutions. We excluded from the analysis all patients without a documented clinical outcome (eg, still hospitalized at the moment of data collection, transferred to another hospital, voluntary discharge) or without complete data.

Derivation of the LOW-HARM Score
The score was constructed based on Fagan's nomogram for Bayes theorem and works as a sequential risk estimation that is modified by the risk factors present in a patient. 18 The pretest probability of death was obtained using the reported prevalence of death by age group. 19  injury was defined as a serum creatinine value >1.5 mg/dL; and leukocytosis as a total count >10,000 cells/μL.
We determined the pretest odds (odds of death by age group) using the following formula (pretest odds = pretest probability/(1-pretest probability)). For this, we used the reported probability of dying by different age group 19  With these data, the calculation for the LOW-HARM score is structured as follows (example in Appendix 1): 1. Pretest odds = pretest probability/(1-pretest probability).

Outcomes
The primary outcome was death during hospitalization.

Analysis
Frequency of each risk factor, mean, and standard deviation for the

Sample size calculation
Mexican official estimations expected at least 10,000, critically ill patients. 26 To ensure a representative sample, according to the formula for estimating samples from finite populations n = N*X/(X + N -1), where X = Zα/22 -*p*(1-p)/MOE, Zα/2 is 1.96, MOE is the margin of error, p is 50% (because the actual p is ignored), and N is the population size, data from 385 patients are required to produce a statistically representative sample with an alpha of 0.05%.

Example of the LOW-HARM score calculation in a hypothetical case
To illustrate how the LOW-HARM score is calculated we consider an 83-year-old patient with hypertension who has been diagnosed with COVID-19 and admitted to the hospital (Figure 1). At admission, he presents with a SpO 2 < 88%, leukocytes >10,000 cells/mm 3 , lymphocytes <0.8 cells/ mm 3 , troponin > 99th percentile, and a serum creatinine < 1.5 mg/dL. Due to his age, this patient's pretest probability of dying is 14.8% (according to Centers for Disease Control and Prevention reports 19 ). This probability is converted to pretest odds (pretest odds = pretest probability/(1-pretest probability) = 0.174). This value is then multiplied by the calculated LR+ for each risk factor to obtain posttest odds (hypertension = 2.06, SpO2 < 88% = 6.85, elevated troponin = 6, leukocyte count > 10 000 cells/μL = 4.23, lymphocyte count < 800 cells/μL = 2.89, serum creatinine > 1.5 mg/dL = 4.23) or by 1 when any of these is absent (in this case, serum creatinine, which was For this hypothetical patient, the posttest probability of death during his hospitalization is 99% (Figure 1). For ease of use, this process is automated in a freely available web app: www.lowharmcalc.com.

Characteristics of study subjects
We obtained data from 438 patients. A total of 38 patients were excluded, leaving 200 patients per group. Their clinical and demographic characteristics are summarized in Table 1. All components of the LOW-HARM score were significantly more frequent in the group of patients who died than in the group of patients who survived their hospitalization ( Table 2).

Predictive performance of the LOW-HARM score
Sensitivity, specificity, positive and negative predictive values, and their corresponding AUCs for different cutoff values are presented in Table 3. The cutoff value of 25 has the highest AUC (0.9); however, it has a specificity (the probability of correctly identifying a survivor with  Using a score cutoff is as useful as the number of times this cutoff is met. In this case, 105/400 (26%) patients had a score above 65, which means it is possible to predict mortality with a specificity of 97.5% and a positive predictive value of 96% in more than a quarter of hospital admissions. On the other hand, it should be considered that mortality in hospitalized patients with COVID-19 has markedly improved because of the refinement of triage systems, the standardization of therapeutic protocols and awareness of early symptoms in the general population. Therefore, end-organ damage at admission is expected to be less frequent.

DISCUSSION
Accurately predicting which patients will not survive hospitalization can guide optimal resource allocation at emergency departments and support clinicians in their decisionmaking process. Additionally, accurate prediction of certain outcomes can help informing patients and their relatives about prognosis.
We present the LOW-HARM score, a novel, easy-to-use, and easy- Having a cutoff value can be useful for decisionmaking. A frequently used method for choosing a cutoff value is to use the value with the largest AUC. In our score, the largest AUC was observed using a cutoff of 25 (0.90, 95% confidence interval: 0.87-0.93). However, because it is possible that clinicians at emergency departments could use the dichotomized version of the score to allocate healthcare resources, we propose a 65-point cutoff value because, in this context, we believe it is preferable to choose a cutoff with a high specificity to correctly identify the highest number of patients that will survive, even if they ultimately die (therefore preserving their "eligibility" for resource allocation), than having a high sensitivity and identifying the highest number of patients that will die, even when they could have survived (therefore denying their "eligibility" for resource allocation).

CONCLUSION
The LOW-HARM score measured at the time of admission has high accuracy in predicting mortality in patients diagnosed with COVID-19 requiring hospitalization. This score provides a disease-specific tool that uses easily obtainable variables making it useful for resourcelimited settings.

ACKNOWLEDGMENTS
To Ricardo Sanginés for developing the web app for the score calculation. All authors thank their respective institutions for their support.

CONFLICT OF INTEREST
None of the authors declares financial interests or personal relationships that could have influenced the work reported in this study.

AUTHOR CONTRIBUTIONS
ASMdesigned the prediction score. JGM, BAMG, and ASR collabo- All authors revised this manuscript and data analysis.