We conducted a single-center retrospective cohort study of consecutive patients with brain metastases (BM) treated at The Jikei University School of Medicine Daisan Hospital between January 2014 and December 2022. Baseline demographic and clinical variables were recorded and are reported in the Results (Table 1); no descriptive counts are repeated in Methods.

“Synchronous” BM (s-BM) was defined as an absolute interval ≤30 days between the first ascertainment of the primary tumor and the first ascertainment of BM, regardless of order; “metachronous” (m-BM) as >30 days. Ascertainment dates were the earliest imaging or histopathology confirmations. (We did not use “after start of primary therapy” as a criterion to avoid conflict with the time-window definition).

Five time-zero variables were prespecified: age, Karnofsky Performance Status (KPS), extracranial metastases (ECM), lung mass, and digestive-tract mass. Clinically interpretable cutoffs were set a priori (age <70 y; KPS ≥70). Because the score is intended for use at initial consultation, when histology may be unavailable, we used imaging-based surrogates for primary site when necessary: “lung mass” (discrete intrathoracic mass on chest radiograph/CT judged by the treating team to represent a presumptive lung primary) and “digestive mass” (discrete mass in the esophagus, stomach, intestine, pancreas, or hepatobiliary system on CT/US). These definitions used the earliest imaging before brain-directed therapy and were not retroactively altered by later pathology to avoid look-ahead/incorporation bias.

Variables significant in univariable Cox received 1 point; those retaining significance in a multivariable Cox model received 2 points, yielding a 0–7 total. As sensitivity analyses, we (i) derived a coefficient-weighted score from the multivariable β estimates and (ii) fitted a reduced score including only multivariable-significant factors; both were compared with the simple scheme using C-index and AIC (Supplement).

For external benchmarks, we implemented the original descriptions of RPA (1), GPA (2), SIR (3), and BSBM (4), using their published category definitions (canonical sources cited).

The primary endpoint was overall survival (OS) from initial consultation to death; patients alive at last contact were censored.

Kaplan–Meier curves were compared with two-sided log-rank tests. Prognostic effects were estimated using Cox proportional-hazards models. Proportional-hazards (PH) assumptions were assessed by the Grambsch–Therneau test on Schoenfeld residuals (cox.zph) for global and variable-specific tests; diagnostic plots were inspected. To compare systems beyond hypothesis testing, we reported Harrell’s C-index (discrimination), Akaike information criterion (AIC) (parsimony), and time-dependent AUCs at 6 and 12 months (timeROC); for categorical systems (RPA, GPA, SIR, BSBM), discrimination was evaluated via the Cox linear predictor. Bootstrap internal validation (B = 500 resamples) provided optimism-corrected C-indices. Missing data were handled by complete-case analysis per model. All tests were two-sided with α = 0.05. Analyses were performed in R and EZR (based on R) (5).

Median overall survival (mOS) was 5.2 months (95% CI, 4.1–7.3) for s-BM and 6.4 months (95% CI, 3.5–8.0) for m-BM. We did not detect a statistically significant difference between groups (Figure 1); this two-group comparison is exploratory and not powered for equivalence/non-inferiority. Our primary analyses focus on developing and validating the s-BM prognostic score. As pretreatment background, the s-BM group had more extracranial metastasis, more primary lung cancer, and fewer breast cancer cases, and in terms of treatment, surgery was more frequent and chemotherapy—especially conventional regimens—was less frequently administered in s-BM. Cohort characteristics.

We analyzed 297 patients overall (s-BM n=64; m-BM n=233). Key baseline features are summarized in Table 1, including age, sex, primary tumor sites, KPS, extracranial metastases, lesion characteristics, and initial treatments. Values are presented as median (IQR) or n (%) (Table 1).

In s-BM, univariable analyses identified age <70 years, KPS ≥70, absence of extracranial metastasis (ECM), presence of primary lung cancer, and absence of primary digestive cancer as factors associated with prolonged mOS (Table 2). In multivariable analysis including these five factors, KPS ≥70 and absence of ECM remained independently associated with longer mOS (Table 3).

Based on the above findings, we constructed a five-item score (age <70, KPS ≥70, no ECM, lung mass present, no digestive mass). Patients were stratified into three categories: Score A (0 points; n=6), Score B (1–4 points; n=31), and Score C (5–7 points; n=27) (Table 4).

mOS for Score A, Score B, and Score C was 57.2 months (95% CI, 8.3–NA), 7.6 months (95% CI, 4.2–12.8), and 3.2 months (95% CI, 2.1–4.6), respectively (Figure 2A), showing a clear stepwise gradient across categories.

For external context, the cohort was also classified by four established systems. RPA: Class I (n=5), Class II (n=21), Class III (n=38); GPA: 0–1.0 (n=43), 1.5–2.5 (n=20), 3.0 (n=1), 3.5–4.0 (n=0); SIR: 1–3 (n=34), 4–7 (n=30), 8–10 (n=0); BSBM: 0 (n=32), 1 (n=11), 2 (n=12), 3 (n=9).

Corresponding mOS values were as follows: RPA Class I, II, III: 57.2, 7.2, and 3.7 months (95% CI, 15.7–NA; 3.9–10.6; 2.5–6.4) (Figure 2B). GPA 0–1.0, 1.5–2.5, 3.0: 3.5, 13.4, and 2.9 months (95% CI, 2.5–6.3; 7.6–28.5; NA–NA) (Figure 2C). SIR 1–3 and 4–7: 6.4 and 6.7 months (95% CI, 2.7–8.0; 3.5–13.9) (Figure 2D). BSBM 0, 1, 2, 3: 3.4, 6.8, 11.2, and 28.5 months (95% CI, 2.3–6.2; 2.5–6.7; 2.7–24.6; 4.2–NA) (Figure 2E, Table 5). Among all systems, only our proposed score separated all categories distinctly across the cohort.

To complement Kaplan–Meier comparisons, we fitted univariable Cox models for each system in s-BM. Our score achieved the highest discrimination (Harrell’s C-index 0.690) and the lowest AIC (375.57) compared with RPA (C = 0.631/AIC=378.65), GPA (0.666/386.13), SIR (0.629/388.63), and BSBM (0.671/379.67). Global Schoenfeld tests indicated no PH violations for any model (our score p=0.983). (Supplementary Table S1) Global Schoenfeld tests did not indicate PH violations for our score (p = 0.983) or for the comparator models (all non-significant), supporting the validity of the Cox analyses. We summarized performance in Table 6 (Harrell’s C-index, AIC, and 6/12-month AUCs). A compact one-panel plot (Figure 3) shows C-index point estimates across systems for visual comparison. In sensitivity analyses, the β-weighted variant achieved similar discrimination but a higher AIC, and the multivariable-only score did not outperform the 5-factor scheme; thus, we retained the pragmatic 1–2 point score for the main analyses (Table 6; Supplementary Tables S1, S2).

Time-dependent AUCs favored our score at both landmarks: 0.779 (6 months) and 0.795 (12 months), versus RPA 0.668/0.667, GPA 0.729/0.706, SIR 0.662/0.702, and BSBM 0.711/0.782 (categorical scores evaluated using Cox linear predictors). (Supplementary Table S1).

Bootstrap internal validation (B = 500) yielded an optimism-corrected C-index of approximately 0.691 for our score, closely matching the apparent estimate, suggesting minimal optimism. Comparator systems had corrected C-indices of ~0.625–0.671.

In a sensitivity analysis, a coefficient-weighted variant of our score (multivariable Cox linear predictor) showed C-index 0.694 and AIC 381.9—comparable discrimination but worse parsimony than the simple scheme. An integer approximation of the weighted model is provided in the Supplement. (Supplementary Tables S1, S2).

This study was conducted at a single center and retrospectively analyzed. Although bootstrap internal validation (B = 500) suggested minimal optimism, the score has not yet undergone external validation. All modeling and performance estimates were derived from the same institution and period, so generalizability across centers, imaging protocols, disease mix, and systemic therapy patterns remains uncertain. Given the modest size of the s-BM cohort (n=64; ~60 events) and the use of the same cohort for both model development and validation, a non-negligible risk of overfitting remains. We prespecified the five predictors and their cut-offs and avoided data-driven stepwise selection to mitigate this risk, and the optimism-corrected C-index (0.691) was close to the apparent estimate; nevertheless, residual optimism cannot be excluded. A limitation of this time-zero design is possible misclassification of the primary site when using imaging surrogates; such an error is likely non-differential and would attenuate associations. External, multicenter validation with standardized imaging adjudication and, where available, pathology cross-check is warranted. In line with reporting guidance for prediction models, the present work represents model development with internal validation; external, multicenter—preferably prospective—validation and potential recalibration are required before broad implementation. We plan such validation in an independent, multicenter s-BM cohort.

Finally, systemic therapy has evolved rapidly in recent years (e.g., immune checkpoint inhibitors and targeted agents). Our development cohort spans both the pre-immunotherapy era and its early adoption. However, the score was based on regimen-agnostic pretreatment factors; its performance may differ in a setting dominated by immunotherapy. Given the limited number of patients treated exclusively with immunotherapy in our s-BM cohort, we did not perform a restricted analysis for this subgroup. This limitation will be addressed in future external, multicenter validation. Treatment-era heterogeneity remains essential, and we plan to pursue an immunotherapy-only subanalysis in a contemporary, multicenter cohort.