Clinicopathologic characteristics associated with long-term survival in advanced epithelial ovarian cancer: an NRG Oncology/Gynecologic Oncology Group ancillary data study

Abstract

Objective

To identify clinicopathologic factors associated with 10-year overall survival in epithelial ovarian cancer (EOC) and primary peritoneal cancer (PPC), and to develop a predictive model identifying long-term survivors.

Methods

Demographic, surgical, and clinicopathologic data were abstracted from GOG 182 records. The association between clinical variables and long-term survival (LTS) (>10 years) was assessed using multivariable regression analysis. Bootstrap methods were used to develop predictive models from known prognostic clinical factors and predictive accuracy was quantified using optimism-adjusted area under the receiver operating characteristic curve (AUC).

Results

The analysis dataset included 3,010 evaluable patients, of whom 195 survived greater than ten years. These patients were more likely to have better performance status, endometrioid histology, stage III (rather than stage IV) disease, absence of ascites, less extensive preoperative disease distribution, microscopic disease residual following cyoreduction (R0), and decreased complexity of surgery (p<0.01). Multivariable regression analysis revealed that lower CA-125 levels, absence of ascites, stage, and R0 were significant independent predictors of LTS. A predictive model created using these variables had an AUC=0.729, which outperformed any of the individual predictors.

Conclusions

The absence of ascites, a low CA-125, stage, and R0 at the time of cytoreduction are factors associated with LTS when controlling for other confounders. An extensively annotated clinicopathologic prediction model for LTS fell short of clinical utility suggesting that prognostic molecular profiles are needed to better predict which patients are likely to be long-term survivors.

INTRODUCTION

Studies by the Gynecologic Oncology Group (GOG) and others describe multiple clinical and pathologic characteristics that are associated with recurrence and survival in advanced epithelial ovarian cancer (EOC) patients [1–8]. In a retrospective analysis of a large group of EOC patients treated with primary cytoreduction and intravenous platinum and paclitaxel chemotherapy on one of six GOG trials, Winter and colleagues confirmed that increasing age, impaired performance status, mucinous or clear-cell histology, and gross residual disease were independent predictors of decreased progression free (PFS) and overall survival (OS). Particularly noteworthy was the impact of residual disease volume in both stage III [6] and stage IV [7] disease. Patients with complete gross resection had a significant survival advantage over those patients with any volume of gross residual. Further studies from the GOG also revealed that patients who presented with diffuse intraperitoneal disease had worse outcomes compared to those with limited intraperitoneal or retroperitoneal spread even after resection to microscopic residual (R0) [2, 5].

The majority of advanced stage EOC recurs within the first two years following diagnosis and over half of deaths occur within five years of diagnosis. The Society of Gynecologic Oncology recommends subspecialty follow-up for patients every three months for the first two years following primary treatment and every 6 months for another three years [9]. This 5-year milestone, after which many patients return to their referring providers, may limit understanding of long-term outcomes and awareness of patients who experience late recurrences. Our recent analyses of GOG-182 revealed that progression and recurrence continued beyond 5 years in patients who had optimal cytoreduction, finally reaching a plateau approximately 7 years after diagnosis [3–5]. In addition, updated outcomes from GOG-111 for patients with sub-optimal cytoreduction show a continual decline in cancer related survival until at least 10 years (written correspondence with W.P. McGuire). These data mirror Sopik and colleagues’ report in which five-year case-fatality rates decreased by 7.5% form 1973-1999, but the 12-year rates decreased only 1.2% suggesting improvements in survival at five-years have not significantly impacted cancer specific mortality when viewed over a longer time horizon[10].

Clinical features, biomarkers, and environmental factors that predict survival past ten years are not known and it remains unclear if the absence of poor prognostic factors or presence of positive characteristics, alone or in combination, predicts long-term survival or cure. Understanding what predicts ovarian cancer long-term survival offers the opportunity to provide improved personalized care. Physicians could direct patients likely to have long-term survival to conventional treatments, while those at higher risk could be channeled to trials of novel therapies. Such selection may enhance the ability of clinical trials to demonstrate clinical benefit. The purpose of this study was to identify clinical and pathologic factors associated with long-term survival and to determine the accuracy of a prediction model using these variables.

METHODS

Eligible patients for this analysis were United States participants in the Gynecologic Cancer InterGroup (GCIG) evaluation of platinum-based treatment regimens in advanced-stage EOC and primary peritoneal cancer (GOG182-ICON5). Patients with stage III or IV disease underwent primary cytoreductive surgery before randomization to 1 of 5 platinum and paclitaxel-based chemotherapy regimens. The randomized study found no statistically significant treatment differences in PFS or OS [11].

Extensive demographic, clinical, and pathologic data were abstracted from study charts and databases. Operative notes and pathology reports were individually scrutinized to compute preoperative disease distribution scores (DS) [2, 4] and surgical complexity scores (CS) [12] as previously described. Briefly, preoperative disease score was defined as (1) DS-low, with pelvic and retroperitoneal spread; (2) DS-moderate (DS-mod), with additional spread to the abdomen, but sparing the upper abdomen; or (3) DS-high, with the presence of upper abdominal disease affecting the diaphragm, spleen, liver, pancreas, or porta-hepatis. Surgical complexity score was based on the complexity and number of surgical procedures. Patients were classified into groups based on the total CS: CS-low (score 1-3), CS-mod (score 4-7), or CS-high (score >8). Patients cytoreduced to R0 were compared to those with gross residual ≤1cm and >1cm. OS was defined as the number of months between study enrollment and any-cause death or last follow up.

The GOG182-ICON5 trial enrolled a total of 4,312 patients. 3,699 patients were enrolled by United States cooperative groups and had charts available for detailed abstraction of disease burden and residual disease status [11]. These patients comprised the final sample for this analysis. Patients with time to death or last follow up greater than 120 months were designated as long-term survivors based on the extended survival curves of GOG 182 and prior GOG trials described above. We excluded 689 patients who were alive with less than 120 months of follow-up at the time of analysis leaving 3010 evaluable subjects. The reference group contained patients who died from any cause within 120 months of enrollment. For this analysis specifically, long-term survival at ten years will be annotated as LTS. Baseline characteristics were summarized by outcome status. Differences between the LTS group and the reference cohort were tested using Wilcoxon and the Pearson Chi Square tests as appropriate. Clinically and statistically significant factors were used as covariates to develop models to predict LTS. Multivariable logistic regression models were used to quantify associations between baseline factors and LTS status. Predictive accuracy of the models was described using receiver operating characteristic (ROC) curves.

Using the entire dataset, bootstrap methods were used to specify a prediction model for LTS [13]. Backward selection of the full model was performed on 2000 bootstrapped datasets using a retention threshold of p<0.20. Main effects were retained to support significant interaction terms. The final model was specified to include covariates retained in at least 70% of the bootstrap samples. Predictive accuracy of the final model was assessed using AUC in which a random classifier would have an AUC of 0.5, while a perfect classifier equals one, or correctly stratifies 100% of the time. Optimism in the AUC estimates was assessed using methods described by Efron and Harrell [14, 15]. Specifically, coefficient estimates for the final model were obtained from bootstrap samples and applied to the original sample. The resulting AUC estimates gave bootstrap estimates of the prospective AUC and its 95% prediction interval. An AUC = 0.8 is considered to have some predictive utility, and the authors determined that this would be a minimum threshold to warrant validation as a potentially actionable predictor in this disease setting.

RESULTS

Demographic and clinicopathologic variables were compared between the 195 cancer patients with LTS and those with shorter survival. Patients with LTS were more likely to have better performance status, endometrioid histology, stage III (rather than stage IV) disease, absence of ascites, decreased surgical complexity, less extensive preoperative and postoperative disease and lower pretreatment CA-125(Table 1).

Table 1.

Comparison of Demographic and Clinicopathologic Factors Associated with Long-term Survival

		OS > 10	Died <10	Overall	P Value
Overall Count	N	195 (7)	2,815 (94)	3,010 (100%)
Age(years)	Median (Q1 - Q3)	58 (51 – 66)	60 (52 – 67)	59 (52 – 67)	0.10
Race	White	180 (92%)	2,582 (92%)	2,762 (92%)	0.87
	Black	8 (4%)	128 (5%)	136 (5%)
	Asian	2 (0.1%)	47 (2%)	49 (2%)
	Other	5 (3%)	58 (2%)	63 (2%)
BMI	Under	10 (5%)	96 (3%)	106 (4%)	0.25
	Normal	87 (44.6%)	1,132 (40.2%)	1,219 (40.5%)
	Over	50 (25.6%)	796 (28.3%)	846 (28.1%)
	Obese	37 (19.0%)	669 (23.8%)	706 (23.5%)
	Missing	11 (5.6%)	122 (4.3%)	133 (4.4%)
Performance Status	Asymptomatic	112 (57%)	1,293 (46%)	1,405 (47%)	0.02
	Ambulatory	73 (37%)	1,302 (46%)	1,375 (46%)
	in bed < 50%	10 (5%)	217 (8%)	227 (8%)
	in bed > 50%	0 (0%)	3 (0.1%)	3 (0.1%)
Site	Ovary	175 (90%)	2,416 (86%)	2,591 (86%)	0.13
	Other	20 (10%)	399 (14%)	419 (14%)
Histology	Adeno Unspecified	3 (2%)	27 (1%)	30 (1%)	<.01
	Clear	5 (3%)	93 (3%)	98 (3%)
	Endometrioid	19 (10%)	110 (4%)	129 (4%)
	Mucinous	1 (1%)	45 (2%)	46 (2%)
	Mixed	13 (7%)	146 (5%)	159 (5%)
	Undifferentiated	2 (1.0%)	39 (1%)	41 (1%)
	Transitional	3 (2%)	10 (0.4%)	13 (0.4%)
	Serous	149 (76%)	2,345 (83%)	2,494 (83%)
Grade	1	8 (5%)	81 (4%)	89 (3%)	0.11
	2	45 (28%)	499 (22%)	544 (22%)
	3	110 (68%)	1,732 (75%)	1,842 (74%)
Stage	3	185 (95%)	2,353 (84%)	2,538 (84%)	<.01
	4	10 (5%)	462 (16%)	472 (16%)
Ascites	Yes	116 (60%)	2,267 (80%)	2,383 (79%)	<.01
	No	78 (40%)	545 (19%)	623 (22%)
Surgical Complexity Score	CS-Low	31 (169%)	685 (24%)	716 (24%)	<0.01
	CS-Mod	144 (74%)	1,749 (62%)	1,893 (63%)
	CS-High	20 (10%)	381 (14%)	401 (13%)
Residual Disease	R0	89 (46%)	490 (17%)	579 (19%)	<.01
	≤1 cm	76 (39%)	1,448 (52%)	1,524 (51%)
	>1 cm	30 (15%)	873 (31%)	903 (30%)
Preoperative Disease Score	DS-Low	23 (12%)	112 (4%)	135 (5%)	<.01
	DS-Mod	74 (38%)	625 (22%)	699 (23%)
	DS-High	98 (50%)	2,078 (74%)	2,176 (72%)
Pretreatment CA-125	Median (Q1 – Q3)	107 (50 – 271)	232 (100 – 595)	220 (93 – 568)	<.01

The starting bootstrap model was specified to include main effects for age, performance status, ascites, preoperative disease score, surgical complexity, residual disease, stage, pretreatment CA-125, and all second-order interactions (a total of 36 covariates including 28 interaction terms). Each of the main effect terms was statistically significant in ≥ 93% of the bootstrapped samples. None of the interaction terms were retained in at least 70% of the samples. A second model was created consisting entirely of main effects. When this multivariable model was tested, age, performance status and disease score were not statistically significant. However, patients attaining R0 were significantly more likely to achieve LTS than those with gross residual. The surgical complexity of the primary surgery was also significant. Patients presenting with ascites, Stage IV disease, or elevated CA-125 were significantly less likely to achieve LTS. The final predictive model contained the seven main effect terms as shown in Table 2. Disease score was dropped from the model due to its high correlation with surgical effort. Although not significant, age and performance status are natural predictors of survival and had retention rates of 94% and 98% respectively in bootstrap models. In accordance with the statistical analysis plan which called for inclusion of covariates retained in at least 70% of bootstrap models, they were retained in the final model. Full-reduced model chi square tests showed no loss of fit between the full (36 variables) and main effects (7 variables) models (p=0.67)

Table 2.

Covariate Effects in the Final Multivariable Predictive Model OR (95% CI) is the estimated odds ratio for surviving at least 10 years. OR>1 indicates improved odds of LTS relative to the comparison group.

Covariate	P value	Comparison	OR (95% CI)
Age	0.50	10 year increase	0.95 (0.83 to 1.10)
Ascites	0.02	Yes reference: No	0.66 (0.47 to 0.94)
log(CA-125)	<0.01	Continuous	0.78 (0.69 to 0.89)
Residual Disease	<0.01	R0 reference: (≤1cm + >1cm)	2.54 (1.80 to 3.58)
Performance Status	0.20	Asymptomatic reference: Symptomatic	0.81 (0.59 to 1.11)
Stage	<0.01	4 ref: 3	0.35 (0.17 to 0.73)
Surgical Complexity	0.03	CS-High reference: CS-Mod	0.70 (0.42 to 1.16)
		CS-Low vs CS-Mod	0.59 (0.38 to 0.91)

Predictive accuracy of the final model was assessed using the AUC. When fit to the data, the final model had AUC=0.729, as shown in Figure 1. The optimism-adjusted estimate was AUC = 0.724 (95% CI: 0.715 to 0.729). The other ROC curves in Figure 1 illustrate the influence of CA-125, residual disease and ascites on the predictive accuracy of the final model. Accuracy was strongly influenced by CA-125 and residual disease status. Although surgical complexity met the threshold for significance, its inclusion did not improve the ability to identify long-term survivors (AUC 0.729 versus 0.725). Given the performance of the exploratory prediction model, further diminished after optimism adjustment, which fell short of our actionable threshold, we did not pursue validation of these data.

Figure 1. ROC Curve of the final multivariate model compared to univariate predictors.

ROC curves of the final predictive model using 7 main effects, compared to predictive ability of pre-treatment Ca-125, residual disease after surgery and presence of ascites.

We performed a similar exploratory analysis as described above limiting the dataset to high grade serous histology acknowledging that the molecular and clinical biology may differ significantly from that of low grade serous and non-serous histologies. The backward selection model was retained but the loss of samples left this model under-powered for interaction terms. This did not materially affect the overall model fit and the AUC for long-term survivors in the high grade serous patients was statistically equivalent to the AUC in the full dataset.

DISCUSSION

Recent studies highlight a number pathologic and clinical variables associated with long-term survival among patients with EOC [16–21], but these associations are inconsistent across studies and details regarding distribution of disease, complexity of surgery, volume of residual disease and types of adjuvant therapies are often unavailable for analysis. In the current study, all patients were treated similarly on a clinical trial and this extensively annotated data revealed significant associations between LTS and performance status, histology, stage, presence of ascites, preoperative extent of disease, postoperative residual disease and surgical complexity in univariable analyses.

Population-based studies consistently associate long-term survival with age and in some, race [16–18]. These were not associated with LTS in our analysis, with age lacking independent significance perhaps due to its collinearity with performance status. Similarly, standardized treatment and follow-up may have limited the impact of race. Although our data did not identify an association of LTS with tumor grade, serous and clear cell histology as well as stage IV disease negatively impacted LTS. In addition, patients with ascites were less likely to have LTS. Recent investigations highlighted that distribution of disease is an important predictor of residual disease as well as a determinant of PFS and OS [2, 3]. Our current data showed that long-term survivors are less likely to have widespread disease involving the upper abdomen (DS-high) and more likely to have R0 at the completion of surgical cytoreduction. Surprisingly though, approximately 15% of patients that survived greater than 10 years had suboptimal residual disease after primary cytoreduction.

Preoperative CA-125 was noted to be much lower among LTS patients (median 107, Q1 50 – Q3 271) compared to women with less than 10-year survival (median 232, Q1 100 – Q3 595). Limited data (serial CA-125s was not required in the original protocol) prevented an assessment of time to normalization and correlations with LTS. A previous meta-analysis by the GOG inclusive of protocols 111, 114, 132, 152, 158, 162, and 172 found pre-treatment CA-125 was associated with PFS and OS [22]. Higher pretreatment CA-125 levels were associated with greater risk for disease progression. Until levels >1,000, there was a 71% increased risk of disease progression compared to normal levels. Although our data was limited by lack of routine reporting of patient BRCA mutational status, recent data indicates that while BRCA status appears to impact platinum sensitivity and short-term survival, it is not associated with long-term survival [20, 23].

Multivariable modelling in our study revealed that a lower CA-125, absence of ascites, stage, and R0 cytoreduction remained significant predictors of LTS when controlling for other influences. Unfortunately the model’s predictive ability (AUC of 0.729) was insufficient for clinical utility as defined by the ability to discriminate long-term survivors sufficiently to alter patient management in ways that might ultimately improve outcomes [24]. Researchers from a multi-center research consortium investigating 10-year long-term survival highlighted the lack of predictability of clinical factors finding that 14% of long-term survival patients had suboptimal cytoreduction, 11% had an initial platinum free interval less than 12 months, and 53% had recurrent disease [19]. Though not evaluated in this analysis, successful prediction of long-term survivorship will likely depend on factors beyond currently recognized clinicopathologic characteristics.

Initial studies by Berchuck and colleagues demonstrated unique transcript expression patterns associated with ovarian cancer patients with greater than seven year survival compared to short-term (<3 year) survival [25]. Barlin and colleagues, using their institutional data and The Cancer Genome Atlas identified distinct transcriptional elements in advanced serous EOC patients they considered likely cured by standard therapy. This population was defined by recurrence-free survival of greater than 5 years and was compared to long-term recurrent patients who survived greater than five years [26]. In the largest and potentially most provocative study of its kind to date, Riester and colleagues published a gene expression signature predicting survival in advanced-stage serous ovarian cancer. These investigators created the new gene signatures using a meta-analytic development approach consisting of gene expression most associated with overall survival across major public datasets. The validated novel signatures consistently outperformed other gene signatures to date and took important steps toward clinical utility [24]. As GOG 182 and other more recent reports imply, further impacting survival by altering the route, timing, and dose of the platinum and taxane backbone is unlikely [11,27,28]. Integration of clinicopathologic factors with these novel molecular signatures in the context of newly identified molecular subtypes and proteogenomic analyses may be a logical next step. We believe such composite predictive modeling and biomarkers hold promise to achieve significant impact [29–31].

In addition to limitations noted previously, it is important to emphasize that this analysis was based on the retrospective review of a single prospective study to abstract key data elements and that the use of a single trial ultimately limited power in some subset analyses. Additionally, the study was accomplished during a time of transition to more aggressive surgery with a goal of achieving complete gross resection. Finally, tissue acquisition for translational research was not required by the protocol and therefore we were unable to integrate molecular characteristics into this investigation. Participating institutions were encouraged to submit tissue and blood from protocol patients through a separate GOG banking protocol. A long term survival consortium was recently established and this group identified tissues associated with a subset of patients on GOG 182 as well as other trials. This consortium is actively collaborating to identify promising biomarkers for future reports.

Predicting which patients are likely, or not likely, to have long-term survival with conventional treatments would have the benefit of tailoring clinical trials and novel therapies to high risk and lower risk populations. Such stratification could be a particularly useful adjunct to novel trial designs such as biomarker enriched or stratified trials as well as randomized umbrella trials. By selecting out a high-risk subgroup, clinical trials may better identify clinically relevant advances. Additionally, patients who are likely to do well with conventional therapy could avoid lengthy consolidation or maintenance therapy or the investment of time, energy, logistical inconveniences, or perhaps invasive procedures associated with some clinical trials and current surveillance algorithms. We believe that personalized treatment of patients at low and high risk for recurrence in this way is economically sound given the restricted resources available for funding large multi-institutional trials.

In conclusion, we were unable to achieve an AUC threshold warranting further development of this model into a clinically useful calculator or nomogram. Despite this, it does represent a benchmark (based on most “knowable” clinic-pathologic data) for development of more accurate predictive models that include molecular features alone or in combination with those we have described.