The prognostic role of systemic inflammatory markers on HIV-infected patients with non-Hodgkin lymphoma, a multicenter cohort study

The primary outcome of this study was all-cause mortality. The cumulative risk of death was determined from the data of NHL diagnosis to the end of the observation period. The censor date for survival analysis was 31st December 2012. The observation period ended either on 31st December 2012, or at last follow-up visit or death, whichever occurred first. Patients lost to follow-up contributed to the time at risk until last visit.

At first, the associations of each prognostic variable with all-cause mortality were tested by univariate and multivariate analysis using Cox proportional hazard models. Gender, age at cancer diagnosis, intravenous drug use, AIDS defining event, HBV and/or HCV co-infection, cART, CD4 cell count, HIV-RNA undetectable were included as covariates. The results were expressed as Hazards Ratios (HRs), their 95% Confidence Intervals (95% CIs), and p-values according to Wald test. Due a few missing data in GPS, mGPS, PLR, PI, and PNI, the Cox proportional hazard regression models were adjusted by stabilized inverse probability weights with missing data in prognostic variables considered missing at random [20]. In order to compare the fitness of the model with and without the inflammatory variables, we used the log-likelihood ratio test.

The Area Under the Receiver Operating Characteristic (ROC) Curve (AUC) was used to assess the predictive accuracy of each prognostic variable. As sensitivity analyses, we assessed the consistency of the prognostic role of inflammatory variables in subjects who underwent cART at NHL diagnosis.

Non-linear relationships between continuous prognostic variables and risk of death were assessed by Cox regression models with spline terms of the prognostic variables [21]. We used the Akaike’s information criterion [22] to assess fitting of models with linear and non-linear terms and to choose the number of spline knots.

Subsequently, we derived a risk score of mortality on the basis of the inflammatory, demographic and clinical variables. To this end, the patients were randomly divided into two subgroups by split-sample technique: (i) derivation subgroup including 70% of subjects, and (ii) validation subgroup, including 30% of subjects. In the derivation sample, a Cox proportional regression model was used to evaluate the association between baseline factors and risk of death, using a cut-off of minimal 10 events per variable for first selection of variables. Gender, age at cancer diagnosis, intravenous drug use, AIDS defining event, HBV and/or HCV co-infection, cART, CD4 cell count, HIV-RNA undetectable and PNI were included. Bootstrapping with 1000 replications was performed to check the stability of variables included in the final models. Linear prediction equations, representing the risk scores, were derived from the final model. For visualization purposes, three risk groups were created according to tertiles of the risk scores distribution in the derivation sample. Cumulative incidence rates of mortality among these three groups were compared by the Kaplan–Meier methods and tested with the log-rank test for trend in the derivation sample. In the derivation and the validation sample the AUC was used to assess the predictive accuracy of the risk scores.

The selection of variables for fitting the most parsimonious models was performed using a backward stepwise procedure, with p = 0.20 for retaining each variable in the model. A graphical check on each regressor did not detect major departures from the proportional hazard assumption of the model.

All the statistical tests were two-sided, assumed a level of significance of 0.05 and were performed using Stata version 12.0 (StataCorp, College Station, TX, USA).

Table 3 describes the distribution of inflammatory markers and the association between each of them and risk of death using Cox regression models. Median time from NHL diagnosis to blood testing was 18 days (interquartile range: 3–53 days).

Score 0 was attributed to the majority of patients for NLR, while for PI and PNI most subjects were categorized as score 1. About half of patients were scored with 1 for GPS and mGPS, and with 0 for PLR. No associations were found between prognostic variables and HIV-RNA copy number in subjects with positive HIV-RNA at NHL diagnosis. GPS, mGPS, PI and PNI were independently associated with risk of death using both univariate and multivariate Cox models. These results were also confirmed when we limited the analyses to patients under cART (data not shown).

The likelihood ratio tests for the global fit of the model, before and after addition of inflammatory variables to the full model, were significant for GPS, mGPS, PI and PNI (p < 0.001).

Non-linear relationships between NLR, PLR, PNI and risk of death were evaluated by using multivariate Cox regression models with a restricted cubic-spline for each prognostic score (Figure 1). A trend of increasing risk of death with increasing PNI was observed, although not statistical significant, whereas no trend with PLR and NLR was found.

The ROC analysis showed that mGPS, GPS, PNI and PI had statistically significant discrimination ability, with AUC ranging from 0.60 to 0.69, with follow-up truncated at both 3 and 5 years (Table 4).

Finally, we computed a risk score for death on the basis of the final Cox model. In the derivation sample (93 subjects with PNI available, 34 of which died), age, occurrence of an AIDS defining event within six months from date of NHL diagnosis, no cART assumption, CD4 cell count lower than 200 cells/mm3 and PNI equal to 1 were significantly associated with fatal outcome in univariate Cox regression analysis (Table 5). However, in multivariate analyses, only occurrence of AIDS defining event within 6 months from NHL diagnosis and PNI were independently associated with mortality, also using bootstrapping technique.

The linear prediction equation for risk of mortality (Rm) derived from the final Cox model was: Rm = 0. 703 *(occurrence of AIDS defining event within six months from NHL diagnosis [yes =1/no = 0]) + 0.794* (PNI). The predictive accuracies of this risk score for mortality, as measured by the AUC, were 0.69 (95% CI 0.58 to 0.81) and 0.69 (0.57 to 0.81) in derivation sample, and 0.69 (0.49 to 0.90) and 0.73 (0.53 to 0.92) in validation sample (47 subjects with PNI available, 12 of which died), at 3 and 5 years of follow-up, respectively. As an application of the score, patients were categorized into three risk groups: patients without AIDS defining event within six months from the date of NHL diagnosis and with PNI > 45, patients with AIDS defining event within six months from the date of NHL diagnosis or with PNI < 45 and patients with AIDS defining event within six months from the date of NHL diagnosis and with PNI < 45. According to the 3 risk score categories, cumulative mortality rates were 20.8% (10.4 to 38.9), 31.8% (18.8 to 50.3) and 64.5% (43.3 to 84.8) in the low-, intermediate- and high-risk group (log rank test p < 0.001) (Figure 2).

As a sensitive analysis, we have also computed a prognostic score including PNI and AIDS events on the whole cohort, finding AUCs for this score of 67.0 (56.3-77.6) and 67.9 (57.8-78.1) at 3 and 5 years, respectively.