Development of a disease-specific graded prognostic assessment index for the management of sarcoma patients with brain metastases (Sarcoma-GPA)

Under the auspices of GSF-GETO, a project involving a multi-institutional retrospective analysis project of sarcoma patients with brain metastases (cerebral or meningeal lesions) was developed (BRAINSARC) [17]. Institutional ethics committee approval was obtained for each centre. The database included patients from 15 French, one Swiss and one Canadian centre. The retrospective data collection was limited to patients managed between January 1992 and March 2012, to ascertain homogeneity in histological diagnosis and classification, namely uniform use of the FNCLCC grading system [18], and to ensure adequate follow-up. The results of this analysis are published elsewhere [17]. Utilizing and enriching this GSF-GETO database, we developed the current project of implementation of the original GPA on sarcoma patients and development of a disease-specific index (Sarcoma-GPA).

Data collection procedures for the BRAINSARC project are described in detail elsewhere [17]. Specifically for the current project, data collection was completed, verified and annotated for the GPA components, notably age at BM diagnosis, Karnofsky performance status (KPS), number of brain lesions and presence of extracranial metastases (ECM), as well as for overall survival (OS). For the development of the disease-specific GPA index, data on ECOG performance status, localization of brain metastasis, time to brain metastasis (TTBM), site of ECM, histological subtype and grade were also collected, verified and annotated. For the histological classification, the 2013 WHO Classification of Tumors of Soft Tissue and Bone was used [19].

Overall survival was estimated from the time of BM diagnosis to the date of death or last follow-up. TTBM was estimated from initial sarcoma diagnosis to the time of BM diagnosis.

For the implementation of the GPA index on our sarcoma cohort, data for each of the four index components were coded according to the original GPA score [8] (Table 1). Each patient was attributed an overall score corresponding to the sum of the scores of individual index components. The GPA index was analyzed in four levels, as per the original description, with group cut-offs of 0–1, 1.5–2.5, 3 and 3.5–4. The GPA scores were subsequently correlated with OS. Survival distributions for individual variables but also for each individual index level compared with all other levels were compared with the log-rank and Mann-Whitney tests using a significance level of 0.001. Overall survival curves for each level of the GPA index were estimated using the Kaplan-Meier method, using the same significance level.

For the development of the Sarcoma-GPA index, cut-offs were decided based on previous GPA indices and on biological sense. Given that the study aim was to identify a meaningful, prognostic way of separating patient subgroups in terms of prognosis, in some instances different variations of cut-offs were attempted in order to identify significant, meaningful cut-offs. Prognostic factors for survival were analyzed by two methods: multivariate Cox regression (MCR) and recursive partitioning analysis (RPA). RPA aided in the identification of best splitting rules amongst prognostic factors. This dual MCR-RPA methodology has been previously shown to be an effective tool in the design of prognostic indices [10, 11, 20]. Prognostic factors found to be significant by either method were used to develop and refine the final Sarcoma-GPA index. Optimal cut-offs for groups were chosen to be consistent with previous disease-specific GPA literature (group cut-offs 0–1, 1.5–2, 2.5–3 and 3.5–4), weighing the significant factors in proportion to the magnitude of the hazard ratio such that 4.0 is the best and 0.0 is the worst [6, 10, 11, 13, 14]. Multivariate analysis was performed using the Cox proportional hazards model. Analyzed variables were age, KPS, ECOG PS, sarcoma type (bone versus soft tissue), localization, tumor size, histological grade and type, time to first metastasis, time to brain metastasis, TTBM, BM lesion number and localisation, presence of and type of ECM at the time of BM diagnosis, and all possible two-way interactions. For hazard ratios, the reference category is defined to have a HR = 1, HR > 1 indicating a higher death rate compared to the reference category. The univariate and multivariate analyses were performed separately for the ECOG PS and KPS variables, as these represent the same clinical characteristic (patient general functional status), in an attempt to identify any clinically pertinent difference in their use within the prognostic score. Since the objective was to develop a prognostic index to guide treatment, no treatment-related variable was analyzed. A forward selection procedure with a cutoff p-value of 0.05 was used to establish the initial model.

For the development of the final model, if individual classes within the investigated variables failed to show statistically significant differences of survival, groupings of multiple levels with similar outcomes were explored. Prognostic factors found to be significant by either MCR or RPA were retained in the final MCR model in order to improve its prognostic ability.

Analysis was performed using the SPSS Statistics version 22 (IBM Corp©, 2013).

The development of the sarcoma-specific index was done in collaboration with the team that described the original and disease-specific GPA indices.

The application of the original GPA score in our sarcoma patient cohort did not allow for validation of its prognostic value. The differences in median OS for each GPA index level were not significant for clear discrimination between each subgroup, especially for the higher-scoring subgroups (Additional file 3: Figure S1). Of the individual index components using score-specific cut-offs, the KPS was the most highly significant, showing the best discrimination amongst component levels (p < 0.001) (Additional file 4: Figure S2).

Different variable cut-offs were individually assessed for significance in regard with overall survival, and were subsequently tested in the under development index. The variables identified as significant in the univariate (age, histology, number of CNS metastases, ECOG PS, KPS, TTBM) and multivariate analysis (histology, number of CNS metastases, ECOG PS, KPS) were individually assessed for the index (Fig. 1, Table 3). RPA analysis results were consistent with the MCR analysis, identifying the number of BMs and the ECOG PS as predictive for survival (Fig. 2).

For the development of the sarcoma-specific GPA, the variables and respective cut-offs identified as significant were tested in different combinations. The best performing split levels and groups, as indicated by both MCR and RPA, lead to the identification of the optimal Sarcoma-GPA index, that included the three variables retained as significant: histology, number of CNS metastases and performance status (Fig. 3). The final index used 4-point cut-offs for the prognostic group levels (scores 0–1, 1.5–2.0, 2.5–3 and 3.5–4), consistent with previously reported GPA scores, with the GPA1 group (score 0–1) having the worst prognosis and the GPA4 group (score 3.5–4) having the better prognosis (Fig. 3). The spit levels chosen for the Sarcoma-GPA index for the histology variables were as follows: group H1 (n = 22; adipocytic tumors, including liposarcoma and myxoid liposarcoma), group H2 (n = 111; smooth muscle tumors including leiomyosarcoma; skeletal muscle tumours including rhabdomyosarcoma; chondro-osseous tumors including osteosarcoma; fibroblastic/myofibroblastic tumors including fibrosarcoma; so-called fibrohistiocytic tumors including pleiomorphic MFH”/ undifferentiated pleiomorphic sarcoma; “vascular tumors, including angiosarcoma; tumors of uncertain differentiation including intimal sarcoma), group H3 (n = 89; tumors of uncertain differentiation, including synovial sarcoma, clear cell sarcoma, epithelioid sarcoma, small round cell tumors, undifferentiated sarcomas, and also malignant peripheral nerve sheath tumor/neurofibrosarcoma and one case of phyllodes tumor/cystosarcoma of the breast) and group H4 [n = 24; predominantly alveolar soft part sarcomas (n = 14) and solitary fibrous tumors (SFT)/hemangiopericytoma (n = 7)], with H4 having the best prognosis and H1 the worst; all individual pairwise comparisons showed statistically significant difference. The split level chosen for ECOG PS were 0–1, 2 and 3–4. For the number of CNS metastases, the split levels found to be more informative were 1, 2–4 and > 4 lesions, differently to the original GPA score.

The log-rank test of the final model and all pairwise comparisons showed a statistically significant difference in median OS between each sarcoma-GPA grouping (p < 0.0001) (Fig. 1c). The addition of the variables age, TTBM and presence of ECM in the under construction indices, assessed at several different split-levels, did not improve their prognostic significant, but rather compromised it (data not shown).

Although both ECOG PS and KPS were individually prognostic (Fig. 2e&f), ECOG PS was found to have a better separation power amongst sub-groups within the final index in comparison to KPS (Fig. 3 vs Additional file 5: Figure S3). The c-index for the original GPA was 0.649, which improved to 0.688 using the Sarcoma-GPA. The ROC curves for the original GPA and Sarcoma GPA corroborated the c-index results (data not shown).

As the Sarcoma-GPA index was designed to provide prognostic information independently of treatment modality, the use of the treatment modalities in our patient cohort was assessed for statistically significant difference. Additional file 2: Table S2 shows the distribution of the different treatment modalities in the individual histology sub-groups as defined above for the Sarcoma-GPA index. No statistically significant difference was observed for any of the treatment modalities between histology groups, with the exception of targeted therapy.

We also assessed whether OS was significantly different in the better prognosis groups (H4 and GPA4) according to the different treatment modalities. The differences were not statistically significant, indicating that the use of different treatment modalities did not significantly influence outcome; for the H4 group the p values were 0.222, 0.386, 0.019, 0.061, 1.00 and 0.37, whilst for the GPA4 group the p values were 0.048, 0.125, 0.048, 0.245, 0.938 and 0.107 for the WBRT, SRS, surgery, systemic chemotherapy, targeted therapy and BSC, respectively (not applicable for the intrathecal chemotherapy, as no patient in the H4 or GPA4 groups received this modality). Significance cut-off for the above was the same used for the construction of the Sarcoma-GPA index, i.e. p < 0.001).