Clinicians use prognostic models and nomograms to inform patients and to aid in making treatment decisions. In prostate cancer, prognostic models have been developed for use in management of men with all stages of prostate cancer, from localized disease to metastatic castration-resistant prostate cancer (mCRPC).[1, 2]
Commonly used prognostic models and nomograms for men with mCRPC have been derived from patients treated in clinical studies.[2-6] Compared with simple prognostic scores, nomograms can be more cumbersome to use in daily practice because of the longer time needed for data entry.
Inflammation is recognized as one of the hallmarks of cancer. The neutrophil-to-lymphocyte ratio (NLR) can be derived easily from the differential blood count and is a marker of host inflammation. The NLR has been demonstrated to have prognostic value in many solid tumors including mCRPC.[8-10] Here we aimed to establish a simple prognostic score for men with mCRPC who were to be treated with docetaxel. We explored the incorporation of the NLR, hypothesizing that this would add prognostic information and improve accuracy.
MATERIALS AND METHODS
Criteria for reporting recommendations for tumor markers in prognostic studies (REMARK) were followed wherever possible.
Sequential men who received at least 1 dose of docetaxel once every 3 weeks for mCRPC between February 2001 and November 2011 at Princess Margaret Cancer Center (PM) in Toronto, Ontario, Canada were reviewed, and their data served as a training cohort to build a prognostic score. The institutional Research Ethics Board approved the study. The validation cohort consisted of patients treated with docetaxel between April 2004 and November 2012 at the Royal Marsden NHS Foundation Trust (RM) in Sutton, Surrey, United Kingdom. These patients provided written informed consent to collect data on institutional research board–approved protocols.
Patients at PM were identified through pharmacy records that listed all patients who had received docetaxel. For patients treated before 2005, the indication for such treatment was not available, and data from the PM Registry were used to define the diagnosis. Patients who received docetaxel for indications other than prostate cancer were excluded, as were men who received weekly treatment (usually because such patients were frail or had significant comorbidities), docetaxel as second-line chemotherapy, or docetaxel for early prostate cancer, that is, in the context of neoadjuvant or adjuvant trials. Clinical data were extracted manually from the electronic patient record (EPR). Laboratory values were obtained electronically from the EPR, and baseline was defined as any measurement occurring up to 28 days before the first dose of docetaxel. Dates of primary diagnosis and death were obtained from the PM Registry, which includes 99% of patients treated in the institution. Data from all sources were linked using the patient's medical record number.
At RM, patients were identified from clinical trial and research databases as having received docetaxel once every 3 weeks for mCRPC. Laboratory and clinical data were collected manually from EPRs.
Descriptive statistics for categorical variables were reported as frequencies and percentages, and continuous variables were reported as medians and ranges. Categorical variables of both cohorts were compared using Fisher's exact test and the Kruskal-Wallis test for 2 and more than 2 variables, respectively. Continuous variables were compared using the Mann-Whitney U test. Survival was calculated from the date of first docetaxel administration to death, with censoring at date of last contact for patients alive at the cutoff date. The Kaplan-Meier method was used to calculate survival estimates, and survival curves were compared using the log rank test. The association of individual variables with survival was assessed using the Cox proportional hazards model. Variables of interest included age at start of treatment, Eastern Cooperative Oncology Group performance status (ECOG PS), site of metastatic spread, Gleason score (< 8 vs 8-10), serum levels of prostate-specific antigen (PSA), PSA doubling time (PSA-dt), lactate dehydrogenase (LDH), alkaline phosphatase (ALP), hemoglobin, albumin, and NLR (the ratio of the absolute blood neutrophil count divided by the absolute blood lymphocyte count). Laboratory values were treated as continuous variables. Variables with nonnormal distribution (eg PSA, LDH, and ALP) were log-transformed.
The method for selection of variables and their cutoffs was defined prospectively. All continuous variables with a P < .05 in univariable Cox proportional hazards analysis were dichotomized using different cutoffs, with at least 5% of patients falling into each category. Subsequent variable reduction was conducted by assessing discriminatory accuracy (the balance between sensitivity and specificity) for each cutoff using the area under the receiver operator characteristics (ROC) curves (c-statistic). The cutoff with the highest c-statistic was selected for further evaluation. ROC curves and associated c-statistics were calculated based on an outcome of death by 15 months, the median survival in the training cohort. Variables whose cutoffs established 2 groups of men showing the highest hazard ratio (HR) between them and the highest c-statistic (ie, the largest area under the ROC curves, thereby indicating maximal discrimination) were selected for further analysis. Multivariable Cox regression with forward stepwise selection was used for primary analysis, and backward stepwise selection was performed as a confirmatory analysis. In situations in which the cutoff with the highest HR was not the same as the cutoff with the highest c-statistic, both cutoffs were chosen for separate multivariable models. Each multivariable model was then compared by its c-statistic, and the model with the highest c-statistic was selected to establish the prognostic score. The final prognostic score assigned one risk point for each variable with a P < .05 in the final multivariable Cox regression model, and the risk points were used to establish 4 prognostic groups. This final score was then tested in the validation cohort. The final model was also compared with the established Halabi nomogram, which included the presence of visceral disease, Gleason score, performance status, PSA, LDH, ALP, and hemoglobin. All analyses were carried out using SPSS version 20 (IBM Corp., Chicago, IL). All statistical tests were 2 sided, and statistical significance was defined as P < .05. No corrections were made for multiple comparisons.
The training and validation cohorts consisted of 357 and 215 patients, respectively. Patient and tumor characteristics of each cohort are presented in Table 1. Compared with men in the validation cohort, men in the training cohort were older (median age, 71 vs 67 years; P < .001), tended to have a worse ECOG PS (P = .052), had more visceral disease, and had higher median LDH, and a higher proportion had hypoalbuminemia. A greater proportion of men in the validation cohort were treated in clinical trials, more received abiraterone acetate, and the median overall survival of the validation cohort was longer (median, 20 vs 15 months; P < .001).
Table 1. Baseline Characteristics
Training Cohort n = 357
Validation Cohort n = 215
Abbreviations: ALP, alkaline phosphatase; BMI, body mass index; IQR, interquartile range; LDH, lactate dehydrogenase; NA, not available; PSA, prostate-specific antigen; PSA-dt, PSA doubling time.
Because of rounding, not all percentages total 100. Missing data are detailed in Supplementary Table S1.
Comparison adjusted for different normal ranges (rates of patients with hypoalbuminemia in the training and validation cohorts were 21% and 12%, respectively [P = .007]).
In univariable analysis higher ECOG PS, presence of liver metastases, lower levels of hemoglobin and albumin, and higher PSA, LDH, ALP, and NLR at baseline were associated with shorter survival of the training cohort (Table 2).
Table 2. Univariable and Multivariable Analyses of the Training Cohort
UVA Continuous Variables
UVA Dichotomized Variables
Percent of Patients
Abbreviations: ECOG PS, Eastern Coopertive Group Performance Status; MVA, multivariable analysis; PCWG2, Prostate Cancer Working Group 2; UVA, univariable analysis.
Age per 10 years
ECOG PS >1
Number of comorbidities
Gleason sum score 8-10
Lymph node metastatic only (PCWG2 type 3)
Bone metastasis (PCWG2 type 4)
Visceral metastasis (PCWG2 type 5)
Hemoglobin <12 g/dL
Albumin <40 g/L
ALP >2.0 × ULN
LDH >1.2 × ULN
PSA >250 ug/L
log(PSA-doubling time [days])
Log(NLR [neutrophils / lymphocytes])
For ECOG PS, hemoglobin, ALP, PSA, and NLR, the cutoffs giving the highest HR between groups did not have the numerically highest c-statistic, and thus both cutoffs (ie, the one with the highest HR and the one with the highest c-statistic based on ROC curves) were taken forward. A total of 16 multivariable analyses were carried out and c-statistics calculated for every model after assignment of 1 risk point per significant variable. C-statistics ranged from 0.78 to 0.80. The highest c-statistic (0.80) was seen when liver metastasis, hemoglobin <120 g/L, ALP >2.0× upper limit of normal (ULN), LDH >1.2× ULN, and NLR >3.0 were included in the model (Table 2). For the final model, variables and effect sizes were similar to those obtained from multivariable analysis using backward elimination of least significant factors. Exclusion of NLR >3 from multivariable analysis or replacing NLR >3 by cutoffs for neutrophils or lymphocytes (or both) resulted in models with lower c-statistics. Excluding the NLR from the final model resulted in a c-statistic of 0.77.
Four groups, with 0, 1, 2, and 3-5 unfavorable prognostic factors, were created for a total of 335 patients (94% of the sample, the remaining 6% having missing data preventing assessment of all variables). Outcomes for these groups are given in Table 3; an example on how to calculate the score is given in the footnote of Table 3. Survival was shorter with more unfavorable prognostic factors (Fig. 1A), and the c-statistic was 0.79 (95% CI, 0.74-0.84; Fig. 2). Using death by months 24 and 36 as outcomes, c-statistics were 0.78 and 0.81, respectively. This compared favorably with the c-statistics calculated with the total points derived from the Halabi score, which were 0.72, 0.74, and 0.70 for the outcomes of death by months 15, 24, and 36, respectively (Fig. 2).
Table 3. Outcome Final Score
Abbreviation: ref, reference.
Example on how to calculate score: 68-year-old man with bone and lymph node metastases from mCRPC about to start treatment with docetaxel. Baseline hemoglobin 138 g/L, alkaline phosphatase 420 U/L (normal range, <150 U/L), lactate dehydrogenase 178 U/L (normal range, <220 U/L), absolute neutrophil count 6.7 g/L, absolute lymphocyte count 1.8 g/L. This patient has 2 risk points (alkaline phosphatase >2 times upper limit of normal, and NLR >3 [6.7 divided by 1.8 = 3.72]) and thus falls in the intermediate-risk group.
1-Year overall survival, %
2-Year overall survival, %
3-Year overall survival, %
Overall survival, months
95% Confidence interval
Hazard ratio for death
95% Confidence interval
The final score could be calculated for 210 patients from the validation cohort (98%) and showed good separation of the survival curves (Fig. 1B) and fairly good discriminatory accuracy (c-statistic, 0.66; 95% CI, 0.58-0.73). Outcomes for the 3 risk groups are given in Table 3. Using death by months 24 and 36 as outcomes resulted in c-statistics of 0.67 and 0.72, respectively (Fig. 2), which were lower than for the Halabi model in prediction of early mortality. Excluding patients who received abiraterone prior to docetaxel resulted in a c-statistic of 0.68.
The variables included in the final score showed significant association with overall survival (OS) in the validation cohort in univariable analysis and except for albumin < 40 g/L were also significant in the multivariable model (Table S2).
We present a simple risk score for men with mCRPC that provided good prognostic and discriminatory accuracy in both training and validation cohorts. Using the presence of liver metastases, the biochemical variables ALP and LDH, and hematologic variables hemoglobin and NLR, 4 risk categories were identified. The 2-year OS ranged from 43%, for a score of 0, to 3%, for a score of 3-5. Good discriminatory ability was demonstrated, with a c-statistic in the training and validation cohorts of 0.78 and 0.66, respectively. Given the prognostic role of the NLR in prostate cancer, this readily available biomarker may also be used to stratify patients in randomized clinical trials.
The main strengths of this study included the prospective definition of the process of variable selection, the ease of use of dichotomous variables, the external validation of the findings in an independent data set, and the hypothesis-driven incorporation of the NLR. However, our score appeared to be a less accurate predictor of early mortality than the Halabi nomogram. Despite this, an advantage over the Halabi tool is that our score is easier to use and does not require use of complex computer-generated algorithms. Our model also performed well when different outcome states were selected.
Of note, the validation cohort in our study comprised a markedly different population of patients than the training cohort, consisting largely of men recruited to clinical trials and with better prognostic features. Patients in the validation cohort had a lower prevalence of liver metastasis and better median overall survival, and more were treated with abiraterone than in the training cohort. Despite these differences, the risk score performed fairly well in the validation cohort, and the c-statistic improved after exclusion of patients treated with abiraterone before docetaxel.
The prognostic role of the NLR has been shown in more than 60 studies for many solid tumors, including prostate cancer.[8, 9, 12, 13] In 135 men with mCRPC receiving ketoconazole, NLR >3 was associated with poorer progression-free survival, and we have shown a role of the NLR in predicting response to abiraterone. The mechanisms underlying this association remain poorly understood but likely reflect more aggressive behavior in tumors with greater inflammation. This hypothesis is supported by data with C-reactive protein, a well-recognized marker of inflammation, which has also been shown to be prognostic in several solid tumors. In recent years, numerous prognostic markers have been proposed for prostate cancer. These include serum androgens. serum biomarkers based on quantitative analysis of the phosphatase and tensin homolog conditional knockout mouse proteome, serum biomarkers of bone metabolism, molecular markers in key steroidogenic pathways, and circulating tumors cells. The utility of inflammation-based markers in prostate cancer also appears valid in this setting.
This study has limitations. First, it was a retrospective cohort study with manual extraction of clinical data. This may have led to inaccurate assessment of PS (often not mentioned explicitly and therefore coded as missing) and nonavailability of relevant data (eg, Gleason scores). Furthermore, because of its retrospective nature, this study will have lower internal validity than data from clinical trials, as there will be more uncertainty regarding dates of outcomes of interest and more heterogeneity regarding measurement of variables of interest. Other limitations include differences between the patients in the training and validation cohorts, including that the patients in the training cohort were not exposed to newly approved agents (abiraterone, enzalutamide, sipuleucel-T, radium-223, or cabazitaxel), and our inability to evaluate all reported prognostic factors including the presence of pain at baseline (pain was not recorded systematically, and therefore we were unable to collect this information), type of progression, and circulating tumor cell counts. We also selected dichotomized cutoff points, although this was done systematically to optimize discrimination.
Neutrophil and lymphocyte counts are nonspecific parameters that are influenced by concurrent infections, inflammation, and corticosteroid medication, which we could not account for. Inflammation and specifically an elevated NLR have been demonstrated to predict mortality in cardiovascular diseases including acute coronary syndromes, which are relevant conditions in our patient population. In contrast, patients on concurrent corticosteroids such as prednisone may have had an artificially high NLR.
In summary, we derived a simple prognostic score for men with mCRPC before the start of treatment with docetaxel once every 3 weeks. We demonstrated the improved prognostic information obtained by adding the NLR to known clinical and laboratory prognostic factors.[2-4] The NLR is a promising, readily available, and inexpensive biomarker that adds prognostic information to current risk scores. This prognostic score should be further explored in other groups of men with mCRPC including those starting initial chemotherapy and those receiving other second-line agents of proven value including cabazitaxel, enzalutamide, and radium-223.[23-25]
A. Templeton's work was supported by a grant from the Swiss Cancer Research Foundation.