Prediction of Lymph Node Involvement in Patients with Breast Tumors Measuring 3 5 cm in a Middle-Income Setting: the Role of CancerMath

Transcription

1 DOI /s BRIEF ORIGINAL SCIENTIFIC REPORT Prediction of Lymph Node Involvement in Patients with Breast Tumors Measuring 3 5 cm in a Middle-Income Setting: the Role of CancerMath E. N. Pijnappel N. Bhoo-Pathy J. Suniza M. H. See G. H. Tan C. H. Yip M. Hartman N. A. Taib H. M. Verkooijen Ó Société Internationale de Chirurgie 2014 Abstract Background In settings with limited resources, sentinel lymph node biopsy (SNB) is only offered to breast cancer patients with small tumors and a low a priori risk of axillary metastases. Objective We investigated whether CancerMath, a free online prediction tool for axillary lymph node involvement, is able to identify women at low risk of axillary lymph node metastases in Malaysian women with 3 5 cm tumors, with the aim to offer SNB in a targeted, cost-effective way. Methods Women with non-metastatic breast cancers, measuring 3 5 cm were identified within the University Malaya Medical Centre (UMMC) breast cancer registry. We compared CancerMath-predicted probabilities of lymph node involvement between women with versus without lymph node metastases. The discriminative performance of CancerMath was tested using receiver operating characteristic (ROC) analysis. Results Out of 1,017 patients, 520 (51 %) had axillary involvement. Tumors of women with axillary involvement were more often estrogen-receptor positive, progesterone-receptor positive, and human epidermal growth factor receptor (HER)-2 positive. The mean CancerMath score was higher in women with axillary involvement than in those without (53.5 vs. 51.3, p = 0.001). In terms of discrimination, CancerMath performed poorly, with an area under the ROC curve of (95 % confidence interval CI ). Attempts to optimize the CancerMath model by adding ethnicity and HER2 to the model did not improve discriminatory performance. Conclusion For Malaysian women with tumors measuring 3 5 cm, CancerMath is unable to accurately predict lymph node involvement and is therefore not helpful in the identification of women at low risk of node-positive disease who could benefit from SNB. E. N. Pijnappel H. M. Verkooijen Imaging Division, University Medical Center Utrecht, Utrecht, The Netherlands estherpijnappel@icloud.com N. Bhoo-Pathy (&) Julius Center, Department of Social and Preventive Medicine, Faculty of Medicine, University of Malaya, Lembah Pantai, Kuala Lumpur, Malaysia ovenjjay@gmail.com J. Suniza M. H. See G. H. Tan C. H. Yip N. A. Taib Department of Surgery, Faculty of Medicine, University of Malaya, Kuala Lumpur, Malaysia M. Hartman Saw Swee Hock School of Public Health, National University of Singapore, Singapore, Singapore

2 Introduction Sentinel lymph node biopsy (SNB) is the standard method for staging the axilla of women with early breast cancer. In affluent settings, all women with unifocal tumors measuring B5 cm are offered SNB [1]. In case of a positive sentinel node, surgery or irradiation of the axilla is administered [2]. Due to cost constraints in low- and middle-income countries, SNB is offered only to breast cancer patients with a low probability of axillary lymph node metastases. Here, in contrast to affluent countries, patients with tumors measuring between 3 and 5 cm often undergo immediate axillary dissection, as their probability of lymph node involvement is considered too high. Still, a substantial proportion of women have node-negative disease and are being over treated with axillary lymph node dissection (ALND). In order to prevent unnecessary morbidity in these settings, identification of patients with tumors measuring between 3 and 5 cm and a low risk of lymph node involvement would be useful in order to offer SNB in a more targeted and cost-effective way. CancerMath is a freely available online tool that predicts among others axillary involvement based on patient and tumor characteristics [3] ( Its predictive performance has not been validated in non-western settings. We investigated whether, in Malaysian patients with tumors measuring 3 5 cm, CancerMath is helpful in identifying women with low risk of lymph node involvement who would therefore be suitable candidates for SNB. Material and methods Data for this study were retrieved from the breast cancer registry of the tertiary University Malaya Medical Centre (UMMC), Kuala Lumpur, Malaysia. Its hospital-based breast cancer registry includes all patients with newly diagnosed breast cancer since 1993; [4] the study was approved by the UMMC Medical Ethical Review Board. Out of the 4,766 women included in the registry between 1993 and 2010, we identified all women with nonmetastatic breast cancer, with tumors measuring between 3.1 and 5 cm. Those with in situ cancers and those treated with neo-adjuvant chemotherapy were excluded. The final study population consisted of 1,017 patients. For each patient, prospectively collected data on age, ethnicity (Chinese, Malay, Indian, other) tumor size, estrogen-receptor (ER) and progesterone-receptor (PR) status (positive, negative, unknown), histological type (ductal, lobular, mucinous, other), grade (well, moderately, and poorly differentiated, unknown) and human epidermal growth factor (HER)-2 status (positive, negative, equivocal, unknown) were available from the registry. For each individual patient, we calculated an individual probability of axillary involvement by entering age, tumor size, ER/PR status, histological type, and grade into the Breast Cancer Nodal Status Calculator of CancerMath. We compared mean CancerMath scores between women with versus those without axillary involvement using the independent sample t test. To assess the calibration of CancerMath, we classified patients into deciles based on CancerMath score and compared the predicted and observed proportions of lymph node involvement for each decile with the Hosmer Lemeshow test. Discrimination of CancerMath was evaluated by receiver operating characteristic (ROC) analysis, where the area under the ROC curve (AUC) expresses the proportion of patients correctly classified as with or without lymph node involvement. In an attempt to optimize the predictive performance of CancerMath, we built a logistic regression model including CancerMath score, ethnicity (Chinese, Malay, Indian, other), [5] and HER2 status. Calibration and discrimination of the extended model were evaluated as described above. All statistical analyses were performed using SPSS version 20 (IBM, Armonk, NY, USA). Results A total of 520 (51%) patients with tumors measuring 3 5 cm had axillary lymph node metastases. Compared with lymph node-negative patients, those with positive nodes had less often well differentiated tumors, more often ductal histology, and were more often ER positive. Age, ethnicity, tumor size, and PR and HER2 status were not significantly associated with lymph node involvement. CancerMath scores were higher for women with positive lymph nodes than for those without (53.5 vs. 51.3, p = 0.001). The calibration of CancerMath, i.e. the association between CancerMath-predicted probability of axillary involvement (in deciles) and the actual observed axillary involvement was inconsistent, with CancerMath-predicted probabilities differing from observed probabilities of lymph node involvement in most deciles (Fig. 1, p value Hosmer Lemeshow test = 0.545). In terms of discrimination, CancerMath performed poorly (AUC 0.553, 95 % confidence interval CI ). This poor discrimination was present in practically all subgroups, with the exception of women with lobular cancer (AUC 0.690, 95 % CI ) and HER2-positive tumors (AUC 0.574, 95 % CI ) (Table 1). An attempt to improve model performance by adding Her2 status and ethnicity to the CancerMath model did not result in clinically meaningful improvement (AUC 0.572, 95 % CI ).

3 Table 1 Characteristics of 1,017 Malaysian patients with breast tumors measuring 3 5 cm according to axillary lymph node status and corresponding discriminatory value (AUC) of CancerMath LN negative a LN positive a p value b AUC 95 % CI Overall ethnicity 497 (49) 520 (51) Chinese 333 (67) 336 (65) Malay 77 (16) 99 (19) Indian 79 (16) 74 (14) Other 8 (2) 11 (2) Age at diagnosis, years (12) 68 (13) (30) 162 (31) (32) 149 (29) (26) 141 (27) Tumor size (64) 313 (60) (36) 200 (38) Unknown 0 (0) 7 (1) Estrogen-receptor status Positive 231 (46) 280 (54) Negative 223 (45) 214 (41) Unknown 43 (9) 26 (5) Progesterone-receptor status Positive 176 (35) 217 (41) Negative 211 (42) 206 (40) Unknown 110 (22) 97 (19) Her2 status Positive 131 (26) 157 (30) Negative 221 (44) 235 (45) Equivocal 51 (10) 54 (10) Unknown 94 (19) 74 (14) Histology Ductal 424 (85) 481 (93) Lobular 13 (3) 21 (4) Mucinous 23 (5) 5 (1) Other 37 (7) 13 (3) Grade (7) 23 (4) (33) 239 (46) (35) 192 (37) Unknown 126 (25) 66 (13) AUC area under the receiver operating characteristic curve, CI confidence interval, LN lymph node a Data are presented as n (%) b p values based on valid observations Discussion Local expertise for SNB is scarce in Malaysia; the number of surgeons trained to perform the procedure is limited, and this is coupled with overburdened histopathology services in most public facilities. Further, only a few centers in Malaysia have the facility to carry out the radionuclide method as the gamma probes and radiocolloids are expensive, and most surgeons use only the blue dye method. While it was hypothesized that CancerMath may be useful to prioritize candidates for SNB among South-East Asian breast cancer patients with nonmetastatic breast cancer measuring 3 5 cm, it appears that it is unsuitable as a selection tool in our setting. This is the first study validating CancerMath for axillary involvement.

4 Fig. 1 Lymph node involvement observed and predicted by CancerMath for 1,017 Malaysian women with breast tumors measuring 3 5 cm. LN lymph node Patients with involved lymph nodes (%) LN predicted LN observed CancerMath score (decile) A possible explanation for this observation is that CancerMath was developed in the USA, and may therefore be overfitted to Western populations, with small, well differentiated, ER-positive tumors. This contrasts with our study population, which contained only tumors measuring 3 5 cm. In previous work, we have shown that in a similar Asian population, another online prediction tool, Adjuvant! Online [6], also did not perform well in predicting overall survival. Overall, Adjuvant! Online over-predicted 10-year survival by 7 %. In young women and women of Malay ethnicity, over-prediction was 20 and 15 %, respectively. Ethnicity is associated with breast cancer biology in Asia. For instance, Malay patients are more likely than Chinese to have axillary lymph node metastasis at similar tumor sizes, and more hormone receptor-negative and poorly differentiated tumors [7]. The triple-negative and HER2- positive subtypes also predominate in this region, with Malays and native Sarawakian patients showing a higher preponderance for triple-negative breast cancers than the Chinese; the Malays also have a high incidence of HER2- positive tumors [8]. As more than 50 % of women with tumors measuring 3 5 cm have no axillary metastases, a substantial proportion of women is being over-treated by immediate ALND and unnecessarily exposed to complications like lymph edema, pain, shoulder problems, and infections. By improving our capability to predict axillary involvement, we could identify women with a low risk of axillary involvement, offer them SNB, and reduce unnecessary morbidity from ALND. While SNB is accurate in recognizing nodal metastasis, there is increasing evidence that nodal dissection in case of proven axillary involvement, does not improve (disease-free) survival [2]. It is therefore likely that the axillary surgery (in the form of SNB or axillary dissection) will decrease [2]. In conclusion, this study showed that CancerMath did not perform well in predicting lymph node involvement. A simple attempt to optimize the prediction model by adding information on HER2 status and ethnicity to the Cancer- Math score was unsuccessful, as it did not improve the predictive performance. Therefore, Asian-based research, including patients with larger, more advanced tumors, is needed to develop region-specific, accurate axillary involvement prediction tools to optimize management of Asian breast cancer patients. References 1. Spillane AJ, Brennan ME (2011) Accuracy of sentinel lymph node biopsy in large and multifocal/multicentric breast carcinoma: a systematic review. Eur J Surg Oncol 37: Giuliano AE, Hunt KK, Ballman KV et al (2011) Axillary dissection vs no axillary dissection in women with invasive breast cancer and sentinel node metastasis a randomized clinical trial. JAMA 305: Michaelson J, Chen L, Bush D et al (2011) Improved web-based calculators for predicting breast carcinoma outcomes. Epidemiology 128: Bhoo-Pathy N, Yip CH, Taib NA et al (2011) Breast cancer in a multi-ethnic Asian setting: results from the Singapore-Malaysia hospital-based breast cancer registry. Breast 20: Bhoo-Pathy N, Yip CH, Hartman M et al (2012) Adjuvant! online is overoptimistic in predicting survival of Asian breast cancer patients. Eur J Cancer 48:

5 6. Ravdin P, Cronin K, Howlander N (2007) The decrease in breastcancer incidence in 2003 in the United States. N Engl J Med 356: Bhoo-Pathy N, Hartman M, Yip CH et al (2012) Ethnic differences in survival after breast cancer in South East Asia. PLoS One 7:e Devi CRB, Tang TS, Corbex M (2012) Incidence and risk factors for breast cancer subtypes in three distinct South-East Asian ethnic groups: Chinese, Malay and natives of Sarawak, Malaysia. Int J Cancer 131: