From the OnCovid repository, a retrospective registry of SARS-CoV-2-infected patients with cancer in Europe,14 we aimed to evaluate a panel of inflammation-based indices that are linked with the survival of patients with cancer for their prognostic role in the context of COVID-19 diagnosis.7–9 11 12 For the purpose of this study, the combination of hypoalbuminemia and lymphocytopenia (albumin concentration (g/L)+5×total lymphocyte count (10⁹/L)), previously qualified as a nutritional predictor of outcome (PNI), was redefined as the OIS in view of its qualification as an inflammatory predictor in the context of COVID-19 infection. Between February 27 and June 23, 2020, 1318 patients were consecutively referred from 23 academic centers in the UK (n=539), Spain (n=380), Italy (n=374), Belgium (n=19), and Germany (n=6, (online supplemental table 1)). Eligibility criteria for OnCovid included (1) SARS-CoV-2 infection confirmed by nasopharyngeal swab and subsequent reverse transcriptase PCR (RT-PCR)15 and (2) history of solid or hematological malignancy at any timepoint before or during COVID-19 disease course. Both patients with active malignancies and patients in cancer remission were included in the study. Patients were ≥18 years of age, and anonymized electronic medical record data were entered into the Research Electronic Data Capture (Vanderbilt University) tool hosted by the Medical Statistics Unit in Novara, Italy,16 17 which coordinated database access and curation.

Eligible patients (n=1071) were stratified on the basis of a selected panel of five biomarkers of systemic inflammation (NLR, PLR, OIS, mGPS, and PI) at three timepoints: at the last oncological follow-up prior to SARS-CoV-2 infection (pre-COVID-19), at COVID-19 diagnosis, and, in surviving patients, at the first oncological follow-up after SARS-CoV-2 infection with a documented negative SARS-CoV-2 swab. Calculations for each inflammation-based index are shown in online supplemental table 2. Patients were stratified into good versus poor risk subgroups on the basis of predefined categories for the mGPS and PI, depending on whether they had an abnormal CRP and/or albumin (for mGPS) or white cell count (for PI) at the point of biomarker measurement.11 12 For inflammatory biomarkers characterized by a continuous distribution, good and poor risk groups were identified by the median value of the distribution.

Features of severe COVID-19 such as ARDS and septic shock were defined according to WHO guidelines.3 Furthermore, all comorbid conditions, complications, and key COVID-19 symptoms were determined from WHO guidelines.3 Resolution of SARS-CoV-2 infection was confirmed via negative RT-PCR result following nasopharyngeal swab for surviving patients. Active malignancy was defined by presence of oncological disease according to clinical criteria for the corresponding tumor type. Patients categorized as undergoing active cancer therapy had received anti-cancer therapy within 4 weeks of COVID-19 diagnosis, and patients with no treatment history for cancer were categorized as treatment naïve. To evaluate the prognostic value of inflammation-based indices, we randomly divided eligible patients into training and validation sets. Primary study outcome was mortality. Overall survival (OS) was determined from the date of positive SARS-CoV-2 RT-PCR to the date of death or final follow-up.

Data following a normal distribution are presented as mean with SD, and data differing from normal distribution are presented as median with IQR. The Mann-Whitney U test and the Wilcoxon signed-rank test were used as appropriate to compare the medians of non-normally distributed continuous data. Categorical variables, shown as absolute counts or percentages, were compared using Pearson’s χ² test. Kaplan-Meier estimates with a log-rank test were employed for univariable survival plot analysis. Univariable HRs were calculated through logistic Cox regression, and multivariable HRs were calculated through Cox regression with conditional backward elimination using a removal value of 0.10 and entry value of 0.05. Harrell’s concordance index (C-index) was used to evaluate the predictive ability of inflammation-based indices and the multivariable model. 150 bootstrap samples were used to calculate the 95% CI for each C-index, estimate the optimism due to potential overfitting, and provide optimism-adjusted C-index values with corresponding CIs. A two-tailed p value of §amp;lt;0.05 was considered the threshold for statistical significance. Dataset management and statistical analyses were performed using Statistical Product and Service Solutions V.24 for Apple Macintosh OSX (IBM, Armonk, New York, USA). Figures were created in Prism V.8 for Apple Macintosh OSX (GraphPad, San Diego, California, USA).

Patients were involved via retrospective analysis of anonymized patient chart data. Patients did not participate in the planning of this study, and this study involved no interventions.

A total of 1318 patients ≥18 years of age with a history of active malignancy or in cancer remission and possessing a positive COVID-19 diagnosis were consecutively entered into the main OnCovid registry. In total, 247 patients were excluded, leaving a final dataset of 1071 eligible patients (figure 1). Reasons for exclusion were (1) lack of laboratory data to calculate at least one biomarker of interest (n=154), (2) leukemia or myeloma diagnosis (n=98), and/or (3) lack of specification of cancer type (n=2). Patients with leukemia and myeloma were excluded due to the potential confounding effect of the neoplastic derangement on peripheral blood values used to compute bone marrow-derived inflammation indices.18–20 From this final study population (n=1071), patients were randomly allocated into a training set (n=529) or a validation set (n=542) matched for age, comorbid burden, number of COVID-19 complications, and proportion of patients with active malignancy (online supplemental table 3). Out of 1071 eligible patients, the majority were male (n=625, 58.3%) with a mean age of 68.2 years (SD ±13.4, range 21–99 years). Patient distribution across participating academic centers is shown in online supplemental table 1. Genitourinary cancers comprised the most common primary tumor site (n=216, 20.2%), and the majority of patients had evidence of active malignancy (n=687, 64.1%). At COVID-19 diagnosis, 519 (48.5%) patients had localized or locoregional disease; 420 patients (39.2%) had metastatic/advanced disease; and 516 (48.2%) patients were receiving anticancer therapy. Most patients had at least one comorbidity (n=844, 78.8%), most commonly hypertension (n=496, 46.3%) or cardiovascular disease (n=257, 24.0%).

Presenting SARS-CoV-2 symptoms included fever (n=660, 61.6%) and cough (n=542, 50.6%). Of the 957 (89.4%) patients hospitalized, 124 (13.0%) were escalated to intensive or subintensive care. The majority of patients required supplemental oxygen therapy (n=627, 58.5%), and a subset required mechanical ventilation (n=115, 10.7%). Most patients received at least one COVID-19-specific drug treatment (n=708, 66.1%), with the majority receiving an antimalarial or antiviral agent (n=622, 58.1%). A total of 754 (70.4%) patients developed COVID-19 complications, most frequently acute respiratory failure (n=627, 58.5%). At the time of database censoring (June 23, 2020), 375 (35.0%) patients were deceased and 680 (63.5%) were discharged from the hospital or were in-hospital survivors (online supplemental table 4 highlights COVID-19 complications and complete patient demographics).

In eligible patients (n=1071), median NLR, PLR, and OIS at COVID-19 diagnosis were 6.1 (IQR 8.7), 270.1 (IQR 260.6), and 38.5 (IQR 10.5) respectively. In total, 526 (49.1%) patients had NLR≥6, 491 (45.8%) had PLR≥270, and 384 (35.9%) had OIS≤40. mGPS was available in 672 (62.7%) patients at COVID-19 diagnosis, with 503 (47.0%) patients being defined as intermediate (mGPS 1, n=196, 18.3%) or poor risk (mGPS 2, n=307, 28.7%). A total of 828 (77.3%) patients were evaluable for PI, with 686 (64.1%) classifying as intermediate (PI 1, n=546, 51.0%) or poor risk (PI 2, n=140, 13.1%) subgroups.

As shown in online supplemental figure 1, OIS values were significantly lower in patients with active malignancy (median 38.1, IQR 11.2), a factor known to predict for outcome from COVID-19,13 compared with patients in remission (median 39.5, IQR 9.6, p<0.05). Patients with active malignancy were also more likely to classify in poorer mGPS risk groups (p<0.05). A comparison of inflammation scores across tumor types is shown in online supplemental table 5.

Median NLR, PLR, and OIS at COVID-19 diagnosis were studied in relation to patients’ mortality along with the distribution of the mGPS and PI. Mortality from COVID-19 was associated with categorization into the poorer risk group for each index (p§amp;lt;0.05, figure 2A).

To evaluate the relative contribution of active SARS-CoV-2 infection over underlying malignancy in determining the systemic inflammatory response, we evaluated dynamic changes across the pre-COVID-19, at, and post-COVID-19 diagnosis timepoints in a subset of SARS-CoV-2 survivors (n=65, 6.1%). All samples taken at post-COVID-19 diagnosis were defined as measurements taken during the first outpatient oncological follow-up after COVID-19 with a negative SARS-CoV-2 swab at the moment of blood sampling. Patients had data for all indices (NLR, n=55; PLR, n=50; OIS, n=23; mGPS, n=8; and PI, n=13). The mean interval between pre-COVID-19 and at COVID-19 diagnosis timepoints was 270 days (SD±855), and the mean interval between COVID-19 diagnosis and post-COVID-19 diagnosis timepoints was 39 days (SD±24).

As shown in figure 2B, NLR values were significantly higher at COVID-19 diagnosis (median 6.1, IQR 8.7) compared with pre-COVID-19 (median 2.9, IQR 3.1, p§amp;lt;0.001) and post-COVID-19 timepoints (median 2.6, IQR 2.8, p§amp;lt;0.01), whereas pre-COVID and post-COVID-19 values did not differ (p=0.82). Similar dynamic changes were described for the OIS, which appeared lower at COVID-19 diagnosis (median 38.5, IQR 10.5) compared with pre-COVID-19 (median 47.0, IQR 9.6, p§amp;lt;0.001) and post-COVID-19 timepoints (47.8, IQR 10.0, p§amp;lt;0.0001), with no difference noted between pre-COVID-19 and post-COVID-19 measurements (p=0.58). mGPS and PI distribution also varied across timepoints (figure 2B).

Next, we sought to determine the relationship between patients’ inflammatory scores at COVID-19 diagnosis and previously established predictors of mortality or disease severity.13 21–24 As shown in online supplemental table 6, patients in the poor risk groups for all inflammatory indices developed a greater number of COVID-19 complications. Furthermore, patients in poor risk groups for all indices except the PLR were older and had a higher comorbid burden. Patients in poor risk categories for the NLR, mGPS, and PI were less likely to be undergoing anticancer therapy at COVID-19 diagnosis and were more often male (online supplemental table 6).

We also investigated the association between inflammatory markers and laboratory variables known to predict for increased mortality from COVID-19.21–24 As shown in online supplemental figure 2, an NLR of ≥6 was associated with increased lactate dehydrogenase (LDH, p§amp;lt;0.0001), D-dimer (p§amp;lt;0.01), troponin (p<0.001), and ferritin (p<0.01) levels compared with an NLR<6. An OIS≤40 was associated with heightened LDH (p§amp;lt;0.001), D-dimer (p§amp;lt;0.001), and ferritin (p§amp;lt;0.001) levels. Poor risk mGPS and PI categories also exhibited an association with clinical variables.

To accurately evaluate the prognostic performance of inflammation-based indices, we randomly assigned the 1071 eligible patients into a training (n=529) and validation set (n=542, figure 1). Both datasets were matched for key characteristics that are known to influence prognosis from COVID-19, including age ≥65 years, number of comorbidities, number of COVID-19 complications, and presence of active malignancy (online supplemental table 3).13 Table 1 presents complete demographics of both the training and validation sets. Unadjusted mortality rates were similar across datasets (36.1% in the training set and 35.0% in the validation set, p=0.16).

In the training set, all tested inflammatory indices but the PLR were associated with unadjusted mortality rates (p<0.0001), as shown in figure 3A. Analysis of unadjusted mortality in the validation set confirmed these findings (figure 3B). Patients with NLR of ≥6, OIS of ≤40, and poor risk mGPS and PI values had shorter median OS (p§amp;lt;0.001, figure 4 and online supplemental table 7). Additionally, as shown in table 2, patients in poor risk groups for all inflammatory indices except the PLR (p=0.55) presented univariable HRs above 1.00 (p≤0.001). This was confirmed in the validation set analysis (p<0.01; figure 5, table 2 and online supplemental table 7).

In multivariable Cox regression models, the inflammatory indices were evaluated for their independent prognostic ability against established prognostic factors for COVID-19.13 21–24 In the training set analysis, the OIS was the only factor emerging as significantly prognostic with patients in the poor risk group (OIS≤40) presenting a multivariable HR of 1.97 (95% CI 1.19 to 3.26, p=0.008; table 2). The independent prognostic value of the OIS also emerged in multivariable analyses of the validation set (OIS≤40, HR 2.48, 95% CI 1.47 to 4.20, p=0.001; table 2), confirming its external validity as a prognostic factor for COVID-19. Analysis of inflammatory indices for prognostic ability was further evaluated through calculation of Harrell’s C-index (online supplemental table 8). The mGPS (optimism-adjusted C-index 0.634, 95% CI 0.595 to 0.680) and OIS (adjusted C-index 0.603, 95% CI 0.555 to 0.646) emerged as most predictive in the training set. OIS was most predictive in validation set analysis (adjusted C-index 0.611, 95% CI 0.564 to 0.668) followed by the mGPS (adjusted C-index 0.596, 95% CI 0.548 to 0.651; online supplemental table 8).