CC BY-NC-ND 4.0 · Revista Urología Colombiana / Colombian Urology Journal 2020; 29(03): 129-135
DOI: 10.1055/s-0040-1713378
Original Article | Artículo Original
Urologic Oncology/Urología Oncológica

Predicting the Probability of Lymph Node Involvement with Prostate Cancer Nomograms: Can We Trust the Prediction Models?

Predicción de la probabilidad de compromiso ganglionar con nomogramas para cáncer de próstata. ¿Podemos confiar en los modelos de predicción?
1  Division of Urology, Facultad de Medicina, Pontificia Universidad Javeriana, Bogotá, Colombia
,
Catalina Barco-Castillo
2  Department of Urology, Hospital Universitario, Fundación Santa Fe de Bogotá, Bogotá, Colombia
,
Jessica Santander
2  Department of Urology, Hospital Universitario, Fundación Santa Fe de Bogotá, Bogotá, Colombia
,
Laura Zuluaga
2  Department of Urology, Hospital Universitario, Fundación Santa Fe de Bogotá, Bogotá, Colombia
,
Camilo Medina
2  Department of Urology, Hospital Universitario, Fundación Santa Fe de Bogotá, Bogotá, Colombia
,
Carlos Trujillo
2  Department of Urology, Hospital Universitario, Fundación Santa Fe de Bogotá, Bogotá, Colombia
,
Mauricio Plata
2  Department of Urology, Hospital Universitario, Fundación Santa Fe de Bogotá, Bogotá, Colombia
,
Juan Ignacio Caicedo
2  Department of Urology, Hospital Universitario, Fundación Santa Fe de Bogotá, Bogotá, Colombia
› Author Affiliations
 

Abstract

Introduction Prediction of lymph node involvement (LNI) is of paramount importance for patients with prostate cancer (PCa) undergoing radical prostatectomy (RP). Multiple statistical models predicting LNI have been developed to support clinical decision-making regarding the need of extended pelvic lymph node dissection (ePLND). Our aim is to evaluate the prediction ability of the best-performing prediction tools for LNI in PCa in a Latin-American population.

Methods Clinicopathological data of 830 patients with PCa who underwent RP and ePLND between 2007 and 2018 was obtained. Only data from patients who had ≥ 10 lymph nodes (LNs) harvested were included (n = 576 patients). Four prediction models were validated using this cohort: The Memorial Sloan Kettering Cancer Center (MSKCC) web calculator, Briganti v.2017, Yale formula and Partin tables v.2016. The performance of the prediction tools was assessed using the area under the receiver operating characteristic (ROC) curve (AUC).

Results The median age was 61 years old (interquartile range [IQR] 56–66), the median Prostate specific antigen (PSA) was 6,81 ng/mL (IQR 4,8–10,1) and the median of LNs harvested was 17 (IQR 13–23), and LNI was identified in 53 patients (9.3%). Predictions from the 2017 Briganti nomogram AUC (0.85) and the Yale formula AUC (0.85) were the most accurate; MSKCC and 2016 Partin tables AUC were both 0,84.

Conclusion There was no significant difference in the performance of the four validated prediction tools in a Latin-American population compared with the European or North American patients in whom these tools have been validated. Among the 4 models, the Briganti v.2017 and Yale formula yielded the best results, but the AUC overlapped with the other validated models.


#

Resumen

Introducción La predicción del compromiso ganglionar es de suma importancia en pacientes con cáncer de próstata (CaP) que se van a someter a prostatectomía radical (PR). Múltiples modelos estadísticos se han desarrollado para predecir el riesgo de compromiso ganglionar y facilitar las decisiones clínicas de realizar o no linfadenectomía pélvica ampliada (LPA). Nuestro objetivo es evaluar la habilidad de predicción de las mejores herramientas de predicción de compromiso ganglionar en CaP en una población latinoamericana.

Métodos Se evaluaron los datos clínico-patológicos de 830 pacientes con CaP sometidos a PR y LPA entre el 2007–2018. Solo se analizaron os pacientes con 10 o más ganglios extraídos (n = 576). Cuatro modelos de predicción fueron validados en esta cohorte: el modelo de la calculadora online del Memorial Sloan Kettering Cancer Center (MSKCC), el Briganti v.2017, la fórmula de Yale, y tablas de Partin v.2016. Se evaluó el desempeño de los modelos con curvas de características operativas del receptor (COR) y el área bajo la curva (ABC).

Resultados La mediana de edad fue 61 años (rango intercuartílico [RI]: 56–66), mediana de Prostate specific antigen (PSA) 6,81 ng/mL (RI: 4,8–10,1), y mediana de ganglios extraídos 17 (RI: 13–23); se documentó compromiso ganglionar en 53 pacientes (9.3%). La habilidad de predicción del nomograma de Briganti v.2017 ABC (0,85) y la fórmula de Yale ABC (0,85) fueron las más precisas. El modelo del MSKCC y las tablas de Partin v.2016 mostraron AUC de 0,84 ambos.

Conclusiones No encontramos diferencia estadisticamente significativa en el desempeño de los cuatro modelos de predicción validados en esta población latinoamericana comparada con pacientes norteamericanos o europeos en los que estas herramientas fueron desarrolladas. Entre los 4 modelos, el nomograma de Briganti v.2017 y la fórmula de Yale mostraron los mejores resultados; sin embargo, el AUC se sobrepone con los otros modelos validados.


#

Introduction

Prostate cancer (PCa) is the second most frequently diagnosed cancer among males worldwide and in Colombia, being the fifth most common cause of death by cancer among males worldwide and the leading cause of death by cancer in Colombia.[1] [2] Radical prostatectomy (RP) and radiotherapy are the standard of care for patients with PCa and aims to cure the patient, being most effective when the disease is confined to the organ. Currently, 12 to 16% of patients with PCa in the United States of America (USA) are managed with RP.[3] [4] [5] Extended pelvic lymph node dissection (ePLND) remains the most accurate method to detect lymph node (LN) metastases despite its invasive nature and the non-negligible risk of complications.

In the PSA era, the fraction of lymph node involvement (LNI) in localized PCa patients has decreased to < 4%; only 0.87% of low-risk, 2% of intermediate risk and 7.1% of high-risk patients have LNI. Despite ePLND being considered a safe procedure, complications may occur, and their severity may vary considerably. Lymphocele rates range between 22 and 54%, thromboembolic events range between 0 and 8%, ureteral injury in < 1% of ePLND, the most frequent neurologic structure injured is to the obturator nerve with a rate between 0 and 5.1%, which represents a landmark of the ePLND template in retropubic RP versus 1.8% in laparoscopic or robot assisted RP. It is of paramount importance to recognize that both the LN yield and the risk of complications are dependent on the extent of the dissection.[6] [7] [8] [9] [10] [11] [12]

To facilitate the selection of PCa patients who would benefit from ePLND, prediction tools (nomograms, formulas, web calculators) have been designed to predict the probability of LNI preoperatively. These tools have been designed and calculated with logistic regression models, classification and regression trees (CART), artificial neural networks and simple linear formulas. More than 20 models are available to this date to predict LNI in patients with PCa. Some of the models are the nomograms reported by Briganti, 2006, 2006 (# harvested LN), 2007, 2007 (# positive cores), 2012, and 2017, Yonsei and Winter nomograms, linear formulas, such as the Roach formula, the Nguyen Formula and the Yale formula; Partin tables that have been validated multiple times, the last one in 2016, and the prediction models proposed by the MSKCC, Godoy prediction model and their web calculators,[13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] Our aim is to evaluate the prediction ability of the most commonly used and best-performing prediction tools for LNI in PCa (MSKCC web calculator, Briganti v.2017, Yale formula and Partin tables v.2016) in a Latin-American population for the first time.


#

Methods

Patient data were collected retrospectively from our institution PCa database; clinicopathological data of 830 patients with PCa who underwent RP and ePLND between 2007 and 2018 were obtained. Patient anonymity was guaranteed. Internal Review Board (IRB) approval was obtained before collecting data. Only data from patients who had ≥ 10 LNs harvested to guarantee that only data for adequate ePLNDs were collected according to previous validation studies.[14] [16] [19] [27] The following histopathological results were available: PSA, cT-stage (assessed by Digital rectal examination [DRE]), Gleason score, number of biopsy cores, number of harvested LNs and positive LNs. Patients for whom data on biopsy cores taken were missing were excluded from the study. A total of 576 patients fulfilled all of these criteria and were included in the study (n = 576 patients).

The ePLND template applied included harvesting nodes overlying the obturator fossa; the external iliac vessels, the internal iliac artery and the common iliac vessels and the presacral stations may be removed, which would allow for clearing ∼ 75% of all anatomical landing sites of LN metastases.[28] [29] All the LNs harvested were centrally reviewed by a uropathologist at our institution.

Four prediction models were validated, the MSKCC web calculator, (https://www.mskcc.org/nomograms/prostate/pre_op), Briganti v.2017, Yale formula and Partin tables v.2016. Model coefficients were derived and made available at www.evidencio.com for validation purposes. Evidencio is an online platform that allows researchers to translate prediction models into user-friendly online calculators, facilitating the application of prediction models. Cutoff values for each model were 2% (Partin), 7% (Briganti), 20% (MSKCC) and 15% (Yale formula), those were taken from the internal validation study of each prediction model as the best performing cutoff value to predict LNI.[13] [15] [16] [18] [22] [23] [26] [27] [30]

Descriptive statistics were reported in terms of the frequency for categorical variables, the median with interquartile range (IQR) for continuous variables that did not follow a normal distribution. Significant differences (p < 0.05) between groups were assessed using the Fisher exact test for categorical variables and a Mann-Whitney U test for continuous variables in the comparison of positive LN (LN + ) and negative LN (LN-) groups ([Table 1]). Prediction tools performance was assessed using the area under the receiver operating characteristic (ROC) curve (AUC). Logistic regression statistics were used to evaluate models under- and overestimation showing the agreement between predicted and harvested LNI. Characteristics of the logistic regression were described in terms of calibration slope and intercept. The calibration slope reflects whether predicted risks are appropriately scaled with respect to each other over the entire range of predicted probabilities and is ideally equal to 1. The intercept (calibration-in-the-large) is a measure that quantifies whether the average of predictions corresponds to the average outcome frequency and ideally equals 0. The c-statistic represents the probability that individuals with the outcome receive a higher predicted probability than those without. It corresponds to the AUC of the ROC for binary outcomes and can range from 0.5 (no discrimination) to 1.0 (perfect discrimination).[31]

Table 1

Comparison between positive and negative lymph nodes groups and global characteristics

Negative LN n = 513

Positive LN n = 53

Global n = 566

p-value

Preoperative characteristics

 Age (years old)[¥]

61 [56–66]

64 [59–68]

61 [56–66]

0.013

 PSA (ng/mL)[¥]

6.5 [4.79–9.1]

13 [7.1–20]

6.81 [4.8–10.11]

< 0.001

Clinical stage[]

 T1a

1 (0.2)

0 (0)

1 (0.2)

< 0.001

 T1b

9 (1.9)

1 (2)

10 (1.9)

 T1c

262 (54.8)

16 (32)

279 (52.7)

 T2a

142 (29.7)

14 (28)

156 (29.5)

 T2b

44 (9.2)

7 (14)

51 (9.6)

 T2c

8 (1.7)

2 (4)

10 (1.9)

 T3a

8 (1.7)

4 (8)

12 (2.3)

 T3b

4 (0.8)

6 (12)

10 (1.9)

Biopsy

 Total biopsy cores[¥]

13 [12–15]

13 [12–15.25]

13 [12–15]

0.244

 Positive biopsy cores[¥]

1 [0–2]

7 [6–11]

4 [2–7]

< 0.001

 Gleason score[¥]

7 [6.75–7]

8 [7–8]

7 [7–7]

< 0.001

Gleason Grade, Group[]

 1

125 (24.5)

2 (3.8)

127 (22.5)

< 0.001

 2

203 (39.8)

6 (11.3)

210 (37.5)

 3

102 (20)

12 (22.6)

114 (20.2)

 4

68 (13.3)

24 (45.3)

92 (16.3)

 5

9 (1.8)

8 (15.1)

17 (3)

Predicted risk according to the nomograms

 MSKCC nomogram (%)[¥]

4 [2–11.25]

20 [15.5–24.5]

5 [2–14]

< 0.001

 Briganti nomogram (%)[¥]

3 [1–15]

36 [15–58]

3 [1–20]

< 0.001

 Yale formula (%)[¥]

5 [3–8]

15 [10.25–23]

6 [3–10]

< 0.001

 Partin tables (%)[¥]

1 [1–4]

11 [5–13.25]

2 [1–5]

< 0.001

Postoperative findings

 Resected lymph nodes[¥]

17 [13–22.5]

18 [14.5–24.5]

17 [13–23]

0.316

 Positive lymph nodes[¥]

.

2 [1–2.5]

.

.

 pN[]

.

.

53 (9.4)

.

Abbreviations: LN, lymph node; MSKCC, Memorial Sloan Kettering Cancer Center; pN, Pathologic Nodes.


¥ Reported as Median [IQR].


Reported as n (%).



#

Results

Baseline characteristics of the cohort are shown on [Table 1]. A total of 830 patients with PCa who underwent RP and ePLND between 2007 and 2018 at the University Hospital, Fundación Santa Fe de Bogotá are described. Only data from patients who had ≥10 LNs harvested were included; A total of 576 patients were analyzed. The median age was 61 years old (IQR 56–66), the median PSA was 6.81 ng/mL (IQR 4.8–10.1), the median of biopsy cores taken was 13 (IQR 12–15), and the median LN harvested were 17 LNs (IQR 13–23), LNI was identified in 53 patients (9.3%). There were significant statistical differences between patients with LN+ and LN- in PSA value, primary Gleason score, positive biopsy cores and clinical stage.

All prediction models were statistically significant, and their AUCs were between the confidence intervals (CIs). Predictions from the 2017 Briganti nomogram AUC (0,85) and the Yale formula AUC (0,85) were the most accurate; the MSKCC AUC and 2016 Partin tables AUCs were 0.84. Because of overlapping confidence intervals (CIs) for the AUCs, logistic regression models were necessary to predict the best performing models. Receiver operating characteristic plots showing the AUC for the four prediction models are presented in [Fig. 1] and [Table 2].

Zoom Image
Fig. 1 Receiving operator characteristic curve for the four prediction models assessed. MSKCC = Memorial Sloan Kettering Cancer Center.
Table 2

Receiver operating characteristic curves regression

Area

95%CI

p-value

MSKCC nomogram

0.844

(0.764–0.891)

0.032

Briganti nomogram

0.859

(0.764–0.901)

0.035

Yale formula

0.859

(0.804–0.913)

0.028

Partin tables

0.844

(0.785–0.904)

0.03

Abbreviation: MSKCC, Memorial Sloan Kettering Cancer Center.


Calibration plots with logistic regression statistics were used to evaluate models under- and overestimation showing the agreement between predicted and harvested LNI. Calibration plots are shown on [Fig. 2]; Partin tables v.2016 showed a calibration-in-the-large (interception) a = 0.24 and a slope b = 1.16, the Briganti v.2017 nomogram showed an interception a = 0.24 and a slope b = 1.13, the Yale formula showed an interception a = 0.24 and a slope b =1.14, and the MSKCC web calculator showed an interception a = 0.24 and a slope b =1.12; calibration-in-the-large or interception were the same for all prediction models, although the slope varied between prediction models, the closest to 1 were the Briganti v.2017 and the MSKCC web calculator.

Zoom Image
Fig. 2 Logistic regression of the four models with calibration plots. (A) Partin tables v.2016 showed a calibration-in-the-large (Interception) a = 0.24 and a slope b = 1.16. (B) Briganti v.2017 nomogram showed an interception a= 0.24 and a slope b = 1.13. (C) Yale formula showed an interception a= 0.24 and a slope b = 1.14. (D) MSKCC web calculator showed an interception a = 0.24 and a slope b = 1.12. MSKCC = Memorial Sloan Kettering Cancer Center.

#

Discussion

There is still controversy regarding the utility and the ability of most of the prediction models validated to predict LNI in PCa in Latin-American patients; given that these prediction tools have been developed and validated with North-American and European patients with different genotypic and phenotypic traits. Our study is the first of its kind, we externally validated the four most accurate prediction tools in a Latin-American (Colombian) population.[14] [19] The 2017 Briganti nomogram and the Yale formula showed the highest AUC (0.859;, however, the 95% CIs for the other two prediction models overlap with the CIs of the aforementioned prediction tools, hence, it remains uncertain if either of these two models truly predicts LNI better than the other validated models, but solve the question of whether or not these models are able to predict LNI with the same accuracy in Latin-American patients.

Partin was the first to published an LNI prediction tool in 1993, where he used PSA, cT and Gleason score to predict LNI in patients with PCa who underwent RP + PLND. In 2016, the last series of Partin tables were published and it included 4,459 patients treated between 2010 and 2015 with an update of the World Health Organization – International Society of Urological Pathology (WHO-ISUP) classification published in 2016 (Grade, Group) and found an AUC of 0.8 to predict LNI.[23] [24] [25] Briganti was another of the pioneers developing prediction nomograms in PCa patients. He published his first nomogram in 2006, but the difference with the Partin tables was that at the time it was the only nomogram based on the results of ePLND. Gacci et al reported an external validation of Briganti nomograms with a predictive accuracy of 79% with the 2012 model.[32] In 2017, Briganti published his last version of the prediction model including patients treated with open, laparoscopic and laparoscopic-robot-assisted RP; The main advantage of this model is the inclusion of the number of cores taken on the biopsy and the number of positive cores to the prediction variables, confirming that the percentage of positive cores involved with PCa was the most reliable predictor of LNI, the AUC of LNI prediction was 0.876, which overlaps with ours.[15] [22] [26] [27] [30]

Godoy et al, in 2011, published an update of the MSKCC nomogram, they assessed 3,721 patients with PCa managed with RP + ePLND, including only the patients with at least 10 LNs harvested. They found 5.2% of LNI in the whole cohort. The AUC for prediction of LNI was 0.862, resembling our findings.[13] [16] Regarding statistical and mathematical formulas to predict LNI, there has been quite an evolution to finally develop the Yale formula, the latter was developed with data of The National Cancer Institute Surveillance, Epidemiology, and End Results (SEER) database. Only data of patients with at least 10 LNs harvested was included; they found a specificity of 94.9% with a 15% cut-off value.[18] [20] [21]

Two external validations of LNI prediction models in PCa patients have taken place in Bulgarian and Dutch patients. In Bulgaria, Hinev et al externally validated Briganti nomograms from 2006 up to 2012 and the MSKCC prediction tool, developed in 2011 by Godoy et al. They found that Briganti's nomograms showed a higher predictive accuracy as compared with the MSKCC prediction model, reporting a calculated AUC of 0.875 for the 2012 Briganti nomogram, and 0.770 for the MSKCC prediction model.[19] Hueting et al validated in the Netherlands 16 prediction models for LNI in PCa patients.[14] They found that the most recent update of the Briganti model (v2012) and of the MSKCC (v2011) showed the highest AUC (0.76) and (0.75), respectively, which is lower than in the Briganti and Godoy validation studies and in our study. In this study, 27.6% of the patients had LNI, contrary to previous validation studies such as Partin Tables (8%), Godoy-MSKCC (5.2%) and our study, in which the percentage of LNI was 9.4%.[16] [23] [24] [25]

The limitations of our study were that despite the percentage of positive cores involved with PCa was the most reliable predictor of LNI; Godoy and Briganti did not describe a standardized method for taking the biopsy cores. In recent years, Multiparametric magnetic resonance imaging (mpMRI) and fusion-guided biopsies of the prostate are becoming the standard of care for patients with suspicion of PCa.[33] Despite this trend, not all of our patients underwent mpMRI, and external validation of nomograms using data from mpMRI were not feasible.[31] [33] Data from the technique and equipment used in the prostate biopsy was not available in our database, so this could represent a bias.[31] [32] [33] Laparoscopic-robotic-assisted surgery has been increasingly used to perform RP and ePLND in our institution in recent years, mostly since 2016. The methods for harvesting LNs may have been different in the whole series (Open ePLND, Laparoscopic ePLND, Robot-assisted ePLND) and the templates may variate from one technique to the other; we were not able to discriminate the surgical modality in our database.[14] [16] [22] [23] [26] [30]


#

Conclusion

Prediction models for LNI in PCa patients undergoing RP play an important role supporting clinical decision-making. Most of these models have been developed in North American or European patients and have never been validated or assessed in Latin-American patients. We found that the 2017 Briganti nomogram AUC (0,85) and the Yale formula AUC (0,85) were the most accurate prediction models in our population, overlapping CIs for the AUCs of the four models were found and calibration plots showed that Briganti v.2017 nomogram (slope = 1.13) and the MSKCC web calculator (slope = 1.12) were the prediction tools with the lowest prediction variability. We recommend using the Briganti v.2017 nomogram and the MSKCC, given their overlapping AUCs and their lower variability. The present study confirms that PCa prediction models for LNI are accurate and could be used with confidence in Latin-American patients.


#
#

Conflict of Interests

The authors have no conflict of interests to declare.


Address for correspondence

Julian Chavarriaga, MD
Facultad de Medicina, Pontificia Universidad JaverianaBogota
Colombia   

Publication History

Received: 27 January 2020

Accepted: 15 April 2020

Publication Date:
04 August 2020 (online)

© 2020. The Author(s). This is an open access article published by Thieme under the terms of the Creative Commons Attribution-NonDerivative-NonCommercial-License, permitting copying and reproduction so long as the original work is given appropriate credit. Contents may not be used for commercial purposes, or adapted, remixed, transformed or built upon. (https://creativecommons.org/licenses/by-nc-nd/4.0/).

Sociedad Colombiana de Urología. Publicado por Thieme Revinter Publicações Ltda
Rio de Janeiro, Brazil


  
Zoom Image
Fig. 1 Receiving operator characteristic curve for the four prediction models assessed. MSKCC = Memorial Sloan Kettering Cancer Center.
Zoom Image
Fig. 2 Logistic regression of the four models with calibration plots. (A) Partin tables v.2016 showed a calibration-in-the-large (Interception) a = 0.24 and a slope b = 1.16. (B) Briganti v.2017 nomogram showed an interception a= 0.24 and a slope b = 1.13. (C) Yale formula showed an interception a= 0.24 and a slope b = 1.14. (D) MSKCC web calculator showed an interception a = 0.24 and a slope b = 1.12. MSKCC = Memorial Sloan Kettering Cancer Center.