Predicting intracranial progression following stereotactic radiosurgery for brain metastases: Implications for post SRS imaging

Abstract

Purpose

Follow-up imaging after stereotactic radiosurgery (SRS) is crucial to identify salvageable brain metastases (BM) recurrence. As optimal imaging intervals are poorly understood, we sought to build a predictive model for time to intracranial progression.

Methods

Consecutive patients treated with SRS for BM at three institutions from January 1, 2002 to June 30, 2017 were retrospectively reviewed. We developed a model using stepwise regression that identified four prognostic factors and built a predictive nomogram.

Results

We identified 755 patients with primarily non-small cell lung, breast, and melanoma BMs. Factors such as number of BMs, histology, history of prior whole-brain radiation, and time interval from initial cancer diagnosis to metastases were prognostic for intracranial progression. Per our nomogram, risk of intracranial progression by 3 months post-SRS in the high-risk group was 21% compared to 11% in the low-risk group; at 6 months, it was 43% versus 27%.

Conclusion

We present a nomogram estimating time to BM progression following SRS to potentially personalize surveillance imaging.

Introduction

Brain metastases (BM) are a common site of metastatic disease presentation and progression occurring in approximately 20-40% of cancer patients.^1–3 Treatment modalities for BMs include surgery, whole-brain radiation (WBRT), and stereotactic radiosurgery/radiotherapy (SRS/SRT), either alone or in combination.² While historically, WBRT was the mainstay of palliative treatment for unresectable BMs, recent data outlining the potential for post-WBRT neurocognitive decline has resulted in an increasing use of SRS in patients with limited intracranial (IC) disease.^3–5 SRS provides good treated-tumor control⁶ and less treatment-induced neurocognitive side effects in both the intact and post-operative setting, and, thus, has become a de facto standard of care for patients with three or fewer lesions.^5,7,8

Randomized evidence suggests that the omission of WBRT correlates with a lower rate of treated-metastasis control and increased rate of new brain metastasis.^5,7 While the differences in treated-tumor control and the development of new brain metastases do not translate into differences in survival or time with functional independence, treatment with SRS requires close serial magnetic resonance imaging (MRI) monitoring to detect new brain metastases while they are small, asymptomatic, and amenable to an appropriate salvage therapy. Such a strategy appears to offer a reduced risk of adverse effects,^3–5,7–9 better outcomes and lower cost.¹⁰ However, BM patients have variable survival,^12,13 and guidelines¹¹ do not routinely take into consideration individual patient factors influencing the time to development of a new brain metastasis.

Therefore, we performed a multi-institutional retrospective analysis exploring factors influencing intracranial progression in SRS-treated patients. We hypothesized that patient and disease-specific criteria would better define time to intracranial progression which in turn could personalize the post-SRS surveillance imaging schedule, potentially reduce health care costs, and improve confidence in predicting and, subsequently, treating early asymptomatic disease progression.

Methods

Patients

We identified consecutive patients treated with SRS for brain metastases at three institutions between January 1, 2002 to June 30, 2017. Eligible patients had baseline cranial MRI imaging to confirm presence of brain metastases and at least one follow-up cranial imaging scan after completion of radiosurgery. Patients were at least 18 years of age and all histologies were included. Prior surgical resection and/or WBRT was allowed. Data were collected on institutional IRB approved protocols.

Patient demographic information, performance status (ECOG), histopathologic diagnosis, tumor molecular markers, extent of extracranial and cranial disease, systemic therapy history including timing, agents, and cycles, evidence of extracranial (EC) disease progression, and dates of surgical resection of intracranial metastases and/or WBRT were recorded. BM diagnosis-specific graded prognostic assessment (DS-GPA, a prognostic score for overall survival derived from factors such as age, performance scores, mutational status, etc.) was calculated using published guidelines.^12,14–17 Dates of initial cancer diagnosis, BM development, and EC metastatic disease diagnoses were retrieved from radiology and pathology records. Radiosurgery treatment characteristics such as number, location, and volume of intracranial (IC) tumors as well as the radiation fractionation and doses delivered (Gy) were also recorded.

Patient Follow-up and Salvage Therapy

Patient charts and IC imaging were reviewed for the presence or absence of progression and symptoms at each follow-up visit. Intracranial progression was classified as distant (untreated brain progression), treated metastases (TM), or dural/leptomeningeal. TM relapse was confirmed by either biopsy or clinical and imaging features strongly suggestive of recurrence within the prior treatment field. Treatment data related to management of progressive intracranial disease involving any combination of SRS, surgery, WBRT, or palliative care were captured.

Study Endpoints

Overall survival (OS) was calculated from SRS completion to either the date of death or last follow-up for living patients. Survival data, including the date of death, were obtained from patient records, national social security database and cancer registries. Time to either asymptomatic or symptomatic progression was evaluated from the SRS completion date to date of imaging showing IC progression.

Statistical Analysis

The Kaplan-Meier method was used to estimate OS and progression-free survival (PFS). To build the nomogram for risk of IC progression, patients were randomly assigned to training and testing sets in a 2:1 ratio, stratifying on primary tumor site and time to initial metastases from cancer diagnosis (>5 years vs ≤ 5 years). Sex, age, race, performance status (ECOG), prior chemotherapy history (Y/N), surgical resection history (Y/N), WBRT history (Y/N), cancer histology, presence of EC metastases, number of treated BMs (1, 2, 3+), time from cancer diagnosis to any initial metastasis, and total BM volume were considered candidate variables for the prognostic model. Genetic markers such as EGFR, ALK, and BRAF mutational status were of interest, but available for less than 20% of the sample and subsequently not included.

Missing data were imputed using multiple imputation with 25 copies and model building was performed on each imputation. Variables were selected using the stepwise method on proportional hazard models and based on statistical significance, using a threshold of 0.15. Since imputation introduced variability in the proposed model, the frequency of selection was used to determine the final model used to build the nomogram. All variables selected for at least 70% (or n=18) of the interim models were included in the final model.

Relative (scaled) point values were assigned for the nomogram using the estimated regression coefficient values. The total scores were calculated in the training set. Distribution of the scores, discriminant ability, and consistency in the training and testing sets were considered when deciding cutpoints. Model accuracy was evaluated using the time-dependent AUC. Cumulative probability of intracranial progression at the selected times were derived from Kaplan-Meier estimates of survival. Statistical analyses were performed in SAS 9.4 and R 3.4.3.

Results

Patient Characteristics

We identified 755 evaluable patients treated with SRS for BMs during the 15-year study period. Patient and tumor characteristics are presented in Table 1. Median patient age at time of SRS was 60 years (range 22-91 yrs) and 41.2% percent were men. The vast majority (88.0%) had good (ECOG 0-1) performance status. The most common tumor subtypes were non-small cell lung cancer (44.6%), breast (19.5%), skin/melanoma (17.1%), and renal cell (7.2%). DS-GPA scores ranged from 0-4 with a median of 2.5. Receptor status was known for 96.6% of breast cancer cases and was comprised of estrogen receptor (ER)/progesterone (PR)+ HER2+ (21.1%), ER/PR- HER2+ (22.5%), ER/PR+ HER2- (29.3%), and ER/PR- HER2- (23.8%). Mutational analyses were performed as clinically indicated, typically in the later years of our cohort with EGFR mutations present in 37 patients, ALK mutations in 14 patients, and BRAF mutations in 36 patients.

Table 1.

Patient characteristics.

Characteristics	N (%)
Patient sample size (N)	755
Age (y)
Median (range)	60 (22-91)
Not recorded	22 (2.9)
Sex
Female	441 (58.4)
Male	311 (41.2)
Not recorded	3 (0.4)
Race
White	566 (75.0)
African American	104 (13.8)
Other	13 (1.7)
Unreported	72 (9.5)
Performance Status (ECOG)
0	323 (42.8)
1	341 (45.2)
2+	27 (3.6)
Not recorded	64 (8.5)
DS-GPA
Median (range)	2.5 (0-4)
Missing†	110 (14.5)
Primary Tumor Site
Non-small cell lung	337 (44.6)
Breast	147 (19.5)
Skin/Melanoma	129 (17.1)
Renal	54 (7.2)
Other	85 (11.3)
Unknown	3 (0.4)
Extracranial disease
Yes	501 (66.4)
Controlled	150 (19.9)
Uncontrolled	309 (40.9)
Unknown	42 (5.6)
No	254 (33.6)
Time to initial metastases from cancer diagnosis (days)
Extracranial
Median (range)	91 (0-8879)
Unknown	1 (0.1)
Intracranial
Median (range)	442 (0-8878)
Unknown	6 (0.8)

Abbreviations: ECOG, Eastern Cooperative Oncology Group; DS-GPA, diagnosis-specific graded prognostic assessment.

^†Data needed to calculate DS-GPA, such as performance status and extracranial disease stability, were missing for these patients.

Initial metastases (EC or IC) developed a median of 152 days (range 0-8879 days) after initial cancer diagnosis, while median time to initial IC metastases was 442 days (range 0-8878 days) (Table 1). At time of SRS, EC disease was absent in 254 patients (33.6%). Median time to initial EC metastases for those who developed it was 91 days (range 0-8879 days). Of those with EC disease, 29.9% were stable, 61.7% uncontrolled, and 8.4% unknown. As shown in Table 2, 176 (23.3%) underwent surgical resection and 282 (37.4%) received WBRT for their IC disease prior to SRS. The majority (n=513, 67.9%) received prior chemotherapy as well. Median number of BMs treated with SRS was 1 (range 1-9).

Table 2.

Treatment history.

Characteristic	N (%)
Prior surgical resection
Yes	176 (23.3)
No	579 (76.7)
Prior whole brain radiation
Yes	282 (37.4)
No	473 (62.6)
Prior chemotherapy
Yes	513 (67.9)
No	227 (30.1)
Unknown	15 (2.0)
Number of intracranial metastases treated with SRS
Total	1407
Median (range)	1 (1-9)
Stereotactic radiation fractionation
Single-fraction	1297 (92.2)
Multi-fraction	103 (7.3)
2-fraction	1
3-fraction	21
5- fraction	81
Unknown	7 (0.50)
Stereotactic radiation dose (Gy)
Single fraction
Median (range)	18 (5-25)
Multi-fraction
Median (range)	25 (12-35)

Overall Disease Outcomes

Median OS was 12.5 months (95% confidence interval [CI], 11.5 – 14.2mos) (Figure S1). Four-hundred fifty patients (62.1%) experienced IC progression of which 34.9% (n= 157) presented symptomatically and 65.1% (n= 293) asymptomatically. Cumulative incidence of asymptomatic progression was 22% (95% CI, 19-25%) at 6 months and 38% (95% CI, 34-42%) at 1 year while symptomatic progression was 14% (95% CI, 12-17%) at 6 months and 22% (95% CI, 19-25%) at 12 months (Figure 1). Median OS (12.4 months (95% CI, 11.2- 14.5) in the training set and (12.9 months (95% CI, 11.2- 14.6) in the testing set (log rank p=0.50)) and PFS (9.2 months (95% CI, 7.8- 10.3) in the training set and (8.0 months (95% CI, 6.4- 10.3) in the testing set (log rank p = 0.57)) did not vary between the training and testing sets.

Figure 1.

Cumulative incidence curves of asymptomatic (1a) and symptomatic (1b) intracranial recurrence show overall higher rates of asymptomatic progression following stereotactic radiosurgery.

Factors Associated with Intracranial Progression-Free Survival

When completed, the predictive model for intracranial progression included tumor histology (melanoma vs not), number of treated BMs (1, 2, 3+), history of WBRT, time from cancer diagnosis to initial metastases (EC or IC), and the interaction of histology by number of treated BMs. The HR for patients with no history of WBRT was 1.22 (95% CI, 0.94- 1.58) compared to patients with WBRT (Table 3). The HR for patients who were diagnosed with metastases within 5 years of initial diagnosis was 1.78 (95% CI, 1.17- 2.70) compared to those who were diagnosed more than 5 years from initial diagnosis. The relationship between number of treated brain metastases and PFS differed by histology.

Table 3.

Multivariate progression-free survival analyses.

Variables	Hazard Ratio (95% CI)	P Value
Prior WBRT: no vs yes	1.22 (0.94- 1.58)	0.13
Time to first metastases from cancer dx: within 5 yrs vs >5 yrs	1.78 (1.17- 2.70)	0.0068
Histology
Melanoma
No. of treated BM: 1 vs 2	1.67 (0.77- 3.64)	0.20
No. of treated BM: 1 vs 3+	0.44 (0.24- 0.80)	0.0068
No. of treated BM: 2 vs 3+	0.26 (0.11- 0.61)	0.0019
Not melanoma
No. of treated BM: 1 vs 2	0.62 (0.45- 0.86)	0.0042
No. of treated BM: 1 vs 3+	0.48 (0.35- 0.66)	< 0.0001
No. of treated BM: 2 vs 3+	0.77 (0.54- 1.10)	0.14

Abbreviations: CI, confidence interval; WBRT, whole-brain radiation therapy; dx, diagnosis; BM, brain metastases.

Although PFS differed by institution before any covariate-adjustment, fitting a shared frailty version of the final model with institution as the random effect revealed no differences in frailty estimates by institution and nearly identical fixed effect estimates. Similarly, stratifying the final Cox model on institution resulted in nearly identical estimates.

Predictive ability of the prognostic model was evaluated using the integrated time-dependent AUC, which was 0.58. As a reference, the integrated tAUC of the DS-GPA was evaluated for its OS predictive ability on the training set and was 0.63.

Prognostic Nomogram for Time to Intracranial Relapse

The clinical nomogram developed from the prognostic model is shown in Figure 2. Patients were ultimately dichotomized into high risk (86-160 points) and low risk (0-85 points) cohorts with the cumulative percent chance of IC progression at various time points calculated for each risk group. The model was then validated on the remaining 1/3 patients in the testing set. One-hundred fifteen (46%) and 134 (54%) patients were categorized as high and low risk with median PFS times of 6.4 months (95% CI, 5.3- 10.0) and 9.8 months (95% CI, 6.6- 14.9), respectively (log rank p = 0.002) (Figure S2). The integrated tAUC for the model using the suggested categorizations was 0.57.

Figure 2.

Clinical nomogram for prediction of time to intracranial progression following stereotactic radiosurgery. Total points accumulated from each prognostic variable stratify patients into high and low-risk cohorts, allowing for estimation of intracranial progression probability at various timepoints in follow-up.

Discussion

IC disease progression after SRS can profoundly impact quality of life, neurocognitive outcomes, and overall survival.¹⁰ Patients with BMs comprise a heterogenous population and the timing of new BM development is poorly understood. Therefore, personalized post-SRS imaging intervals are needed to detect BMs prior to symptom onset. In our large, diverse, multi-institutional cohort of patients, we found that 62.1% progressed intracranially after SRS, with a significant proportion having symptomatic progression. We identified prognostic factors of intracranial PFS and created a nomogram to predict individual patient probability of IC progression at various time points. Although several studies have investigated treated-metastasis control and OS following SRS, to the best of our knowledge this is the largest multi-institutional study to build a predictive model of time to IC progression after SRS. Our model provides a simple, useful personalized tool for clinicians and patients to guide appropriate MRI imaging intervals.

Figure S1.

Kaplan-Meier curve of overall survival for patients following stereotactic radiosurgery (SRS/SRT) shows a median survival of 12.5 months.

We found that a greater number of BMs, melanoma histology, and shorter time from cancer diagnosis to initial metastasis (> 5 yrs vs · ≤ 5 yrs) were all associated with higher risk of intracranial PFS independently. Our data largely concur with prior literature as prior studies have reported that age, number of BMs, melanoma histology, performance status, gross tumor volume, prior WBRT, and progressive systemic disease all influence IC progression after SRS.^3,5,18–23 However in addition to these, we identified disease indolence as a factor influencing intracranial progression following SRS.

No tool currently exists to evaluate how different prognostic factors interact to affect an individual patient’s time to IC progression following SRS. Previous prediction models have focused on either OS, an important but separate endpoint that does not aid in patient surveillance, or time to salvage WBRT after SRS, which is of diminishing utility now that patients are increasingly treated for consecutive BMs with further courses of SRS.^{12,14–17,24}

Given the lack of a personalized predictive tool for patients post-SRS, we developed this nomogram to provide anticipatory guidance for both patients and clinicians. Ideally, patients with low risk of progression after SRS could be spared the cost and stress of undergoing surveillance imaging every 3 months. As routine surveillance imaging can induce significant patient anxiety and fear of progression, minimizing imaging frequency among low-risk patients may help alleviate this psychological and financial burden.²⁵ Alternatively, high-risk patients are more likely to rapidly develop new BMs. These high-risk patients may present symptomatically prior to scheduled scans, resulting in increased cost, higher rates of neurosurgical intervention, longer inpatient hospital stays, higher risk of neurological death, and decreased OS.¹⁰ More frequent MRI monitoring in this high-risk cohort would allow for earlier detection of BMs at a time when they are smaller and asymptomatic, making them more amenable to rapid treatment with SRS. However, increased surveillance imaging may also pose risks such as increased healthcare utilization, a theoretical rise in invasive biopsies to differentiate progression vs. pseudoprogression and patient anxiety. Thus, we hope our prediction tool serves to inform clinical management and patient-provider conversations on customized imaging schedules.

To illustrate the utility of this nomogram, consider a 60-year-old patient diagnosed with lung adenocarcinoma three years ago with no prior history of WBRT. The patient, seen in consideration of SRS for two BMs, would be placed in the high-risk group. Subsequently, the cumulative probability of progression at 3 months after SRS—the timepoint at which the first follow-up MRI is currently performed per NCCN guidelines—is 21.4%. Closer MRI monitoring, such as at 2 months post-treatment when the chance of recurrence is already 12.0% per our nomogram, may be considered to detect occult BMs before clinical presentation. Indeed, similar high-risk patients studied on the Alliance (N0574) randomized trial over a similar time period demonstrated a 6 week recurrence rate of 11%.⁵

Additionally, our model complements the widely-used DS-GPA, used to predict OS after initial diagnosis of BM.¹² The DS-GPA continues to help patients and clinicians gauge overall survival and, therefore, select appropriate therapy at BM diagnosis. Our model helps guide clinical management after this initial presentation by elucidating the likely time to IC progression following SRS for each patient based on their unique pathophysiology and treatment history.

Although our study is comprised of a large, diverse, and multi-institutional cohort of patients, it is limited by its retrospective nature and nonrandomized design. Moreover, the study period spans over 15 years during which time many new technologies and systemic therapy regimens were introduced into practice, affecting OS, PFS, and intracranial tumor control (i.e. later generation agents for ALK-rearranged NSCLC and BRAF-mutated melanoma). Although the study sites currently use high-resolution, thin-slice MRI series, such technology was not available in the early years of this study. Such higher resolution technology may be better able to capture smaller metastases, impacting the detected frequency of recurrence. Our prediction model was validated internally and rigorous statistical methods were used to maximize generalizability. Comprised of patients with diverse tumor histologies and largely good performance status, our cohort is representative of typical SRS patients. However, our results should be validated on an external independent cohort, ideally one derived from prospective randomized data, including large groups of patients treated with immune checkpoint and molecularly targeted therapies, not well represented in our patient cohort.

In conclusion, in this large multisite cohort, we tracked patterns of BM progression both radiographically and clinically to construct a novel predictive model for time to IC relapse. Prognostic factors include number of BMs (1, 2, 3+), histology (melanoma vs not), history of WBRT, and time from initial cancer diagnosis to first metastases (> 5 yrs vs ≤ 5 yrs). Our model seeks to guide individualized surveillance imaging regimens to reduce symptomatic BM progression and its associated quality of life, psychosocial, and financial burden.

Figure S2.

Internal validation of prediction model using training and testing sets in a 2:1 ratio. Significant difference in progression-free survival is maintained between the high and low-risk groups (p=0.002).

Acknowledgments

We would also like to thank Nicholas Gerwal for his contributions to data collection.