Development and internal validation of the CLIP-M score for advanced mucosal malignant melanoma: exploratory prognostic stratification and comparative assessment with the RMH score Authors: Burak Pacaci, Buket Erkan Ozmarasali, Adem Deligonul, Ali Fuat Gurbuz, Melek Karakurt Eryilmaz, Suheda Atas Ipek, Nisanur Sariyar Busery, Emre Yilmaz, Mustafa Alperen Tunc, Ali Kaan Guren, Erkam Kocaaslan, Ahmet Demirel, Firat Akagunduz, Nazim Can Demircan, Osman Kostek Published: 2026 ABSTRACT BackgroundMucosal malignant melanoma (MMM) is a rare, aggressive malignancy with poor prognosis in the advanced setting. Although the Royal Marsden Hospital (RMH) score is widely used in advanced malignancies, its utility in advanced MMM remains insufficiently characterized. We aimed to develop a composite prognostic score and compare its performance with RMH.MethodsWe analyzed 80 patients with advanced MMM (metastatic or unresectable) from a multicenter database. Pretreatment inflammatory indices and clinical variables were evaluated. The CLIP-M score (Clinical and Inflammatory Prognostic Score in Mucosal Melanoma) was derived from independent predictors in multivariable Cox regression and internally validated by bootstrap resampling. Predictive accuracy was compared with RMH using time-dependent AUCs, Harrell’s C-index, and Kaplan–Meier estimates.ResultsDuring a median follow-up of 65.2 months, 63 deaths occurred. SII was selected as the biological component. CLIP-M integrated high SII (≥ 789; derived by 12-month time-dependent ROC analysis), age ≥ 65 years, and non–head & neck primary location. Time-dependent AUCs numerically favored CLIP-M over RMH at 12 months (0.731 vs. 0.613), 24 months (0.824 vs. 0.680), and 36 months (0.863 vs. 0.697), with a nominal 36-month between-model difference that did not remain significant after correction for three time horizons (ΔAUC = 0.166; nominal p = 0.039). Bootstrap internal validation showed moderate discrimination, with apparent and optimism-corrected C-indices of 0.648, and acceptable 12-month calibration. CLIP-M stratified patients into low-risk (score 0–1) and high-risk (score 2–3) groups with divergent median overall survival (24.8 vs. 6.4 months; HR 4.03, 95% CI 2.25–7.22; p < 0.001).ConclusionCLIP-M provided exploratory, clinically interpretable risk stratification in advanced MMM and showed numerically higher discrimination than RMH; however, optimism-corrected performance was moderate and external validation is required before clinical implementation.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12885-026-16194-w. Introduction Mucosal malignant melanoma (MMM) is a rare and biologically distinct subtype of melanoma, accounting for approximately 1.3% of all cases, and is characterized by aggressive behavior, lower tumor mutational burden, and lack of association with ultraviolet radiation exposure [1–3]. Due to its occult anatomical location and rapid progression, a substantial proportion of patients present with or rapidly develop advanced disease (unresectable or metastatic), in which prognosis remains poor, with median overall survival typically ranging between 8 and 12 months [4, 5]. In the advanced setting, accurate risk stratification is essential for treatment planning, counseling patients, and clinical trial enrollment. Although the American Joint Committee on Cancer (AJCC) staging system provides an anatomical framework, it does not adequately reflect the biological heterogeneity of host–tumor interactions in advanced disease [6]. In addition, a mucosal melanoma–specific evidence-based staging system has recently been proposed using large-scale natural-history data [5]. However, staging systems primarily characterize anatomical disease extent and were not used as direct comparators in the present study because our objective was to develop a pretreatment clinicobiological prognostic score for patients with advanced MMM, defined as unresectable or metastatic disease, at the time of systemic therapy. Therefore, we benchmarked CLIP-M against the Royal Marsden Hospital (RMH) score—based on LDH, albumin, and number of metastatic sites—which is widely used for risk stratification in advanced malignancies and early-phase oncology trials [7]. However, its specific utility in MMM remains insufficiently characterized. Given the distinct immunobiology of MMM, prognostic models driven primarily by tumor burden and metabolic markers may be insufficient. In this context, systemic inflammatory markers—particularly the Systemic Immune-Inflammation Index (SII)—have been investigated as potential prognostic indicators across solid tumors, yet their comparative performance against established clinical scores such as RMH or nutrition-based indices like the Glasgow Prognostic Score (GPS) has not been evaluated specifically in advanced MMM [8–11]. SII, calculated from platelet, neutrophil, and lymphocyte counts, integrates three components of the host inflammatory and immune response [9]. Biologically, elevated SII may reflect neutrophil- and platelet-mediated tumor progression together with impaired lymphocyte-dependent antitumor immunity [8, 9]. In melanoma, baseline SII and related immune-inflammatory indices have been evaluated mainly in advanced-stage and immune checkpoint inhibitor–treated cutaneous melanoma cohorts, with some studies reporting poorer survival among patients with higher SII and others showing mixed prognostic performance [12, 13]. In mucosal melanoma, our previous registry-based analysis suggested the prognostic relevance of SII as a standalone biomarker across a broader MMM population [11]. However, whether SII or other inflammatory and nutritional indices can be integrated with clinical variables into a disease-specific composite prognostic model for advanced MMM, using laboratory values defined at the advanced-disease time point and benchmarked against RMH, remains insufficiently characterized. Therefore, there is an unmet need for a disease-specific, composite prognostic model tailored to advanced MMM. In this study, we aimed to derive and internally validate a novel composite score, the CLIP-M score (Clinical and Inflammatory Prognostic Score in Mucosal Melanoma), integrating inflammatory and clinical variables, and comparatively assessing its prognostic performance against the widely used RMH score. Materials and methods Study design and patient population This multicenter retrospective cohort study included adult patients (≥ 18 years) with histologically confirmed advanced MMM treated at six oncology centers in Turkey between 2011 and 2025. The present analysis was conducted using an updated multicenter MMM database that had also supported our previous registry-based report [11]. Unlike the prior study, the present work focused specifically on advanced MMM, used laboratory values defined at the advanced-disease time point, developed a composite prognostic score, benchmarked it against RMH, and performed internal validation. Advanced disease was defined as radiologically confirmed distant metastases (Stage IV) or unresectable locally advanced disease requiring systemic therapy. Patients with unresectable locally advanced disease were analyzed together with metastatic cases due to their comparable aggressive clinical course and systemic treatment approach [4, 5, 14, 15]. Among 111 eligible patients, 31 were excluded because of missing baseline clinical and/or laboratory variables required for calculation of inflammatory indices, resulting in a final analytic cohort of 80 patients (Fig. 1). Exclusion criteria included uveal or cutaneous melanoma, active infection, hematologic malignancy, autoimmune disease, or incomplete baseline data. The study was conducted in accordance with the Declaration of Helsinki and was approved by the Marmara University School of Medicine Ethics Committee, which served as the coordinating institutional review board for this retrospective multicenter study (protocol no. 09.2026.26–0116). The predefined eligibility period covered patients treated between 2011 and 2025, and data abstraction and analysis were initiated after ethics approval was obtained in 2026. Fig. 1Patient selection flow diagram Patient selection flow diagram Data collection and definitions Demographic and clinicopathological variables, including age, sex, Eastern Cooperative Oncology Group Performance Status (ECOG-PS), comorbidity status, primary tumor location, metastatic sites, and first- and second-line systemic therapies, were extracted from medical records. Comorbidity status was recorded as a binary variable (present vs. absent); detailed comorbidity subtypes were not consistently available across centers. Baseline laboratory parameters, including complete blood count, serum albumin, C-reactive protein (CRP), and lactate dehydrogenase (LDH) levels, were defined uniformly as the values obtained closest to advanced-disease diagnosis and before initiation of first-line systemic therapy for advanced disease. For de novo metastatic cases, this corresponded to laboratory values at initial advanced diagnosis; for metachronous relapse, laboratory values at relapse before first-line systemic therapy for advanced disease were used. Molecular testing data, including BRAF, KIT, and NRAS mutation status, were extracted when available; however, testing was not uniformly performed across centers or calendar periods and was therefore not included in model development. Inflammatory and nutritional indices were calculated using the following formulas: SII = (Platelet × Neutrophil) / Lymphocyte [9]NLR (neutrophil-to-lymphocyte ratio) = Neutrophil / Lymphocyte [16]PLR (platelet-to-lymphocyte ratio) = Platelet / Lymphocyte [17]PNI (prognostic nutritional index) = 10 × Albumin (g/dL) + 0.005 × Total Lymphocyte Count (/mm³) [18]LIPI (lung immune prognostic index): Calculated based on derived NLR (dNLR = Absolute Neutrophil Count / [White Blood Cell Count - Absolute Neutrophil Count]; dNLR > 3) and LDH (> upper limit of normal) [19]GPS: Score 0 (CRP ≤ 10 mg/L and Albumin ≥ 3.5 g/dL), Score 1 (CRP > 10 mg/L or Albumin < 3.5 g/dL), and Score 2 (CRP > 10 mg/L and Albumin < 3.5 g/dL) [20]RMH Score: Calculated as described by Arkenau et al. [7], assigning 1 point for each of the following: LDH > upper limit of normal, Albumin < 3.5 g/dL, and > 2 metastatic sites. SII = (Platelet × Neutrophil) / Lymphocyte [9] SII = (Platelet × Neutrophil) / Lymphocyte [9] NLR (neutrophil-to-lymphocyte ratio) = Neutrophil / Lymphocyte [16] NLR (neutrophil-to-lymphocyte ratio) = Neutrophil / Lymphocyte [16] PLR (platelet-to-lymphocyte ratio) = Platelet / Lymphocyte [17] PLR (platelet-to-lymphocyte ratio) = Platelet / Lymphocyte [17] PNI (prognostic nutritional index) = 10 × Albumin (g/dL) + 0.005 × Total Lymphocyte Count (/mm³) [18] PNI (prognostic nutritional index) = 10 × Albumin (g/dL) + 0.005 × Total Lymphocyte Count (/mm³) [18] LIPI (lung immune prognostic index): Calculated based on derived NLR (dNLR = Absolute Neutrophil Count / [White Blood Cell Count - Absolute Neutrophil Count]; dNLR > 3) and LDH (> upper limit of normal) [19] LIPI (lung immune prognostic index): Calculated based on derived NLR (dNLR = Absolute Neutrophil Count / [White Blood Cell Count - Absolute Neutrophil Count]; dNLR > 3) and LDH (> upper limit of normal) [19] GPS: Score 0 (CRP ≤ 10 mg/L and Albumin ≥ 3.5 g/dL), Score 1 (CRP > 10 mg/L or Albumin < 3.5 g/dL), and Score 2 (CRP > 10 mg/L and Albumin < 3.5 g/dL) [20] GPS: Score 0 (CRP ≤ 10 mg/L and Albumin ≥ 3.5 g/dL), Score 1 (CRP > 10 mg/L or Albumin < 3.5 g/dL), and Score 2 (CRP > 10 mg/L and Albumin < 3.5 g/dL) [20] RMH Score: Calculated as described by Arkenau et al. [7], assigning 1 point for each of the following: LDH > upper limit of normal, Albumin < 3.5 g/dL, and > 2 metastatic sites. RMH Score: Calculated as described by Arkenau et al. [7], assigning 1 point for each of the following: LDH > upper limit of normal, Albumin < 3.5 g/dL, and > 2 metastatic sites. Primary tumor sites were categorized as head & neck versus non–head & neck based on reported anatomical, biological, and prognostic differences in mucosal melanoma; the non–head & neck category included gastrointestinal, gynecological, urinary/genitourinary, and rare respiratory mucosal sites [14, 15]. Statistical analysis Overall survival (OS), defined as the time from diagnosis of advanced disease to death or last follow-up, was the primary endpoint. Survival distributions were estimated using the Kaplan–Meier method and compared with the log-rank test; median follow-up was calculated using the reverse Kaplan–Meier method. Model development followed Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) recommendations [21], with an emphasis on parsimony and clinical interpretability. Given the modest sample size and the overlapping components of inflammatory and nutritional indices, candidate biomarkers (SII, NLR, PLR, PNI, LIPI, and GPS) were not entered simultaneously. Instead, each index was evaluated in a separate multivariable Cox model adjusted for the same prespecified core clinical covariate set (age group, gender, ECOG-PS, comorbidity, primary site group, and liver metastasis), and a single representative marker was selected for score development to minimize redundancy and collinearity. Across these models, both SII and GPS showed prognostic relevance. SII was selected for score development because it represented a distinct hematologic immune-inflammatory axis, avoided overlap with albumin-based scores such as GPS and RMH, and could be incorporated into a simple bedside score after dichotomization using a prespecified 12-month time-dependent receiver operating characteristic (ROC) threshold. An optimal SII cut-off was determined using Kaplan–Meier–based time-dependent ROC analysis at the prespecified 12-month horizon by maximizing the Youden index; the derived cut-off was subsequently carried forward for score construction. The CLIP-M score was constructed by integrating dichotomized SII (≥ 789), age ≥ 65 years, and primary tumor location (non–head & neck vs. head & neck), with unit weighting assigned to each component [22]. Equality of the retained model coefficients was assessed using a global Wald test to support the unit-weighting approach. After candidate selection, final Cox models were refitted using the forced-entry method, whereby all selected predictors were entered simultaneously rather than through stepwise elimination. Internal validation was performed using bootstrap resampling (1,000 iterations) to assess optimism and calibration. Model development was performed using a complete-case analytic cohort because the candidate inflammatory and nutritional indices required multiple source laboratory variables. Multiple imputation was not used because missingness affected the source variables required to calculate composite biomarkers, and imputing several derived inflammatory/nutritional indices in a small cohort could introduce additional uncertainty into score development. Because patients excluded for incomplete data were not systematically abstracted beyond eligibility screening, a reliable included-versus-excluded baseline comparison could not be performed. Discriminatory performance was assessed using inverse probability of censoring–weighted (IPCW) time-dependent AUCs at 12, 24, and 36 months, with between-model comparisons conducted accordingly. Time-dependent AUC analyses were performed using the ordinal point scores for CLIP-M and RMH, whereas Kaplan–Meier curves and Cox group comparisons used the prespecified dichotomized risk categories. Model fit was additionally compared using Cox model likelihood-ratio chi-square statistics and Akaike information criterion (AIC), with lower AIC indicating better relative fit. To explore whether SII provided incremental prognostic information beyond RMH, exploratory Cox models comparing RMH alone versus RMH plus dichotomized SII were fitted, with model fit assessed using likelihood-ratio testing and AIC; apparent time-dependent AUCs based on Cox linear predictors were also examined. An exploratory calendar-era analysis was performed to determine whether changes in systemic treatment practice over time introduced potential confounding. Calendar era was defined according to the date of advanced disease diagnosis and dichotomized as 2011–2017 versus 2018–2025. Treatment distribution and CLIP-M risk distribution across calendar eras were compared using exact tests, as appropriate. Era-adjusted Cox regression, CLIP-M × calendar-era interaction analysis, and era-stratified Cox models were performed as exploratory sensitivity analyses. For descriptive clinical interpretability, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy of the dichotomized CLIP-M score were calculated for 12-month mortality after excluding patients censored before 12 months. Exploratory 12-month decision curve analysis was performed to compare the net benefit of CLIP-M and RMH across threshold probabilities for 12-month mortality using Cox model–based predicted risks with censoring adjustment. Because SII showed a right-skewed distribution, ln-transformed SII was evaluated as a sensitivity analysis to reduce the influence of extreme values and assess the robustness of the continuous SII association. The proportional hazards assumption for the final CLIP-M model was assessed using Schoenfeld residuals (Supplementary Figure S4). Descriptive statistics, crosstabulations with exact tests, Kaplan–Meier analyses, and conventional Cox regression models were performed using IBM SPSS Statistics version 27.0 (IBM Corp., Armonk, NY, USA). Time-dependent ROC/AUC analyses, bootstrap internal validation and calibration, Schoenfeld residual testing, and exploratory decision curve analysis were performed using R version 4.5.2 (R Foundation for Statistical Computing, Vienna, Austria). Key R packages included survival version 3.8-6, rms version 8.1-0, timeROC version 0.4, survivalROC version 1.0.3.1, riskRegression version 2025.9.17, prodlim version 2025.4.28, and ggplot2 version 4.0.2. A two-sided p-value < 0.05 was considered statistically significant. Results Patient characteristics The median age at diagnosis was 63.5 years (IQR, 56–73), with 43.8% of patients aged ≥ 65 years. Most patients were male (56.3%) and had an ECOG-PS of 0–1 (87.5%). Primary tumor locations were gastrointestinal (47.5%), head & neck (42.5%), gynecological (5.0%), and other mucosal sites (5.0%). At presentation, 45% of patients had de novo metastatic disease, 45% developed metachronous metastases, and 10% had unresectable locally advanced disease. The lung (43.8%) and liver (31.3%) were the most frequent metastatic sites. Radiotherapy-based local treatment was recorded in a subset of patients: 12 patients (15.0%) received adjuvant radiotherapy, 6 (7.5%) received adjuvant chemoradiotherapy, and 5 (6.3%) received definitive chemoradiotherapy; no patient received definitive radiotherapy alone. Second-line systemic therapy was administered to 39 patients (48.8%), including immunotherapy in 26 patients (32.5%), chemotherapy in 8 (10.0%), and BRAF-directed targeted therapy in 5 (6.3%). Molecular testing was incomplete and heterogeneous across the cohort. BRAF testing was available for 63 patients, of whom 10 (15.9%) were mutation-positive; KIT testing was available for 20 patients, of whom 10 (50.0%) were mutation-positive; and NRAS testing was available for 5 patients, of whom 1 (20.0%) was mutation-positive. After a median follow-up of 65.2 months (95% CI, 56.3–74.2), 63 deaths (78.8%) were observed. Baseline clinicopathological characteristics are detailed in Table 1. Table 1Baseline clinicopathological characteristics of patients with advanced mucosal melanoma (n = 80)Characteristicn (%)Age, years Median (IQR)63.5 (56–73) < 6545 (56.3) ≥ 6535 (43.8)Gender Male45 (56.3) Female35 (43.8)ECOG-PS 041 (51.3) 129 (36.3) 28 (10.0) 32 (2.5)Primary Tumor Location Head & Neck34 (42.5) Gastrointestinal38 (47.5) Gynecological4 (5.0) Other mucosal sites4 (5.0)Disease Presentation De novo Metastatic36 (45.0) Metachronous Metastasis36 (45.0) Unresectable Locally Advanced8 (10.0)Sites of Metastasis Lung35 (43.8) Liver25 (31.3) Bone21 (26.3) Brain7 (8.8) Other23 (28.8)RMH Score Components LDH > Upper Limit of Normal25 (31.3) Albumin < 3.5 g/dL11 (13.8) Metastatic Sites > 28 (10.0)First-line Systemic Therapy Chemotherapy54 (67.5) Immunotherapy20 (25.0) None / Best Supportive Care6 (7.5)Second-line systemic therapy Chemotherapy8 (10.0) Immunotherapy26 (32.5) BRAF-directed targeted therapy5 (6.3) No second-line systemic therapy41 (51.3)Comorbidities Present51 (63.8) Absent29 (36.3)Radiotherapy-based treatment history Adjuvant radiotherapy12 (15.0) Definitive radiotherapy0 (0.0) Adjuvant chemoradiotherapy6 (7.5) Definitive chemoradiotherapy5 (6.3)Molecular testing availability BRAF tested63 (78.8) BRAF mutation positive among tested10/63 (15.9) KIT tested20 (25.0) KIT mutation positive among tested10/20 (50.0) NRAS tested5 (6.3) NRAS mutation positive among tested1/5 (20.0)Head & neck primaries included nasal cavity/sinonasal (n = 17), oral cavity (n = 12), nasopharynx (n = 3), uvula (n = 1), and salivary gland (n = 1). Other mucosal sites included urinary/genitourinary and rare respiratory mucosal primaries. Molecular testing availability percentages were calculated using the total cohort as the denominator, whereas mutation positivity percentages were calculated among tested patients. Molecular testing was not uniformly performed across centers or calendar periods; therefore, mutation status was not incorporated into prognostic modelingAbbreviations: IQR Interquartile Range, ECOG-PS Eastern Cooperative Oncology Group Performance Status, RMH Royal Marsden Hospital, LDH Lactate Dehydrogenase Table 1Baseline clinicopathological characteristics of patients with advanced mucosal melanoma (n = 80)Characteristicn (%)Age, years Median (IQR)63.5 (56–73) < 6545 (56.3) ≥ 6535 (43.8)Gender Male45 (56.3) Female35 (43.8)ECOG-PS 041 (51.3) 129 (36.3) 28 (10.0) 32 (2.5)Primary Tumor Location Head & Neck34 (42.5) Gastrointestinal38 (47.5) Gynecological4 (5.0) Other mucosal sites4 (5.0)Disease Presentation De novo Metastatic36 (45.0) Metachronous Metastasis36 (45.0) Unresectable Locally Advanced8 (10.0)Sites of Metastasis Lung35 (43.8) Liver25 (31.3) Bone21 (26.3) Brain7 (8.8) Other23 (28.8)RMH Score Components LDH > Upper Limit of Normal25 (31.3) Albumin < 3.5 g/dL11 (13.8) Metastatic Sites > 28 (10.0)First-line Systemic Therapy Chemotherapy54 (67.5) Immunotherapy20 (25.0) None / Best Supportive Care6 (7.5)Second-line systemic therapy Chemotherapy8 (10.0) Immunotherapy26 (32.5) BRAF-directed targeted therapy5 (6.3) No second-line systemic therapy41 (51.3)Comorbidities Present51 (63.8) Absent29 (36.3)Radiotherapy-based treatment history Adjuvant radiotherapy12 (15.0) Definitive radiotherapy0 (0.0) Adjuvant chemoradiotherapy6 (7.5) Definitive chemoradiotherapy5 (6.3)Molecular testing availability BRAF tested63 (78.8) BRAF mutation positive among tested10/63 (15.9) KIT tested20 (25.0) KIT mutation positive among tested10/20 (50.0) NRAS tested5 (6.3) NRAS mutation positive among tested1/5 (20.0)Head & neck primaries included nasal cavity/sinonasal (n = 17), oral cavity (n = 12), nasopharynx (n = 3), uvula (n = 1), and salivary gland (n = 1). Other mucosal sites included urinary/genitourinary and rare respiratory mucosal primaries. Molecular testing availability percentages were calculated using the total cohort as the denominator, whereas mutation positivity percentages were calculated among tested patients. Molecular testing was not uniformly performed across centers or calendar periods; therefore, mutation status was not incorporated into prognostic modelingAbbreviations: IQR Interquartile Range, ECOG-PS Eastern Cooperative Oncology Group Performance Status, RMH Royal Marsden Hospital, LDH Lactate Dehydrogenase Baseline clinicopathological characteristics of patients with advanced mucosal melanoma (n = 80) Head & neck primaries included nasal cavity/sinonasal (n = 17), oral cavity (n = 12), nasopharynx (n = 3), uvula (n = 1), and salivary gland (n = 1). Other mucosal sites included urinary/genitourinary and rare respiratory mucosal primaries. Molecular testing availability percentages were calculated using the total cohort as the denominator, whereas mutation positivity percentages were calculated among tested patients. Molecular testing was not uniformly performed across centers or calendar periods; therefore, mutation status was not incorporated into prognostic modeling Abbreviations: IQR Interquartile Range, ECOG-PS Eastern Cooperative Oncology Group Performance Status, RMH Royal Marsden Hospital, LDH Lactate Dehydrogenase Derivation of the CLIP-M score The selected marker, SII, was evaluated together with prespecified clinically relevant covariates (age group, sex, ECOG-PS, comorbidity, primary site group, and liver metastasis) in a multivariable Cox model. To obtain a parsimonious bedside score, backward stepwise selection based on the likelihood ratio test was applied; sex, ECOG-PS, comorbidity, and liver metastasis were not retained. The final CLIP-M score included three predictors: high SII (≥ 789), age ≥ 65 years, and non–head & neck primary site. For transparent reporting, coefficients and hazard ratios shown in Table 2 were estimated from a refitted multivariable Cox model including only these three retained predictors. Because regression coefficients were of similar magnitude, each component was assigned 1 point (total 0–3). Full estimates from the prespecified SII models are provided in Supplementary Table S1. Univariable and core-adjusted Cox regression analyses of all candidate inflammatory and nutritional indices are shown in Supplementary Table S4. In these analyses, SII and GPS retained independent associations with OS, whereas NLR, PLR, PNI, and LIPI did not retain independent associations after adjustment. Table 2Multivariate Cox regression analysis and derivation of the CLIP-M scoreVariableCoefficient (B)SEHR (95% CI)P-valueAssigned PointsHigh SII (≥ 789)0.7420.2672.10 (1.24–3.55)0.0061Age ≥ 65 years0.6220.2611.86 (1.12–3.11)0.0171Non–head & neck primary site0.7910.2762.21 (1.28–3.79)0.0041Variables excluded from the final model included sex, ECOG-PS, comorbidity, and liver metastasis. With 63 events and 3 variables in the final refitted model, the final-model EPV was 21. A global Wald test did not reject equality of the three retained coefficients (χ²=0.208, df = 2, p = 0.901), supporting the unit-weighting approach. Candidate inflammatory indices were not entered simultaneously; instead, each index was evaluated in a separate core-adjusted model to reduce redundancy and collinearity. Nevertheless, although the final refitted model had a nominal EPV of 21, this does not fully capture the optimism introduced by candidate biomarker screening, cut-off selection, and backward selection; therefore, residual overfitting cannot be excluded, and the score should be considered exploratory until externally validatedAbbreviations: SE Standard Error, HR Hazard Ratio, CI Confidence Interval, SII Systemic Immune-Inflammation Index Table 2Multivariate Cox regression analysis and derivation of the CLIP-M scoreVariableCoefficient (B)SEHR (95% CI)P-valueAssigned PointsHigh SII (≥ 789)0.7420.2672.10 (1.24–3.55)0.0061Age ≥ 65 years0.6220.2611.86 (1.12–3.11)0.0171Non–head & neck primary site0.7910.2762.21 (1.28–3.79)0.0041Variables excluded from the final model included sex, ECOG-PS, comorbidity, and liver metastasis. With 63 events and 3 variables in the final refitted model, the final-model EPV was 21. A global Wald test did not reject equality of the three retained coefficients (χ²=0.208, df = 2, p = 0.901), supporting the unit-weighting approach. Candidate inflammatory indices were not entered simultaneously; instead, each index was evaluated in a separate core-adjusted model to reduce redundancy and collinearity. Nevertheless, although the final refitted model had a nominal EPV of 21, this does not fully capture the optimism introduced by candidate biomarker screening, cut-off selection, and backward selection; therefore, residual overfitting cannot be excluded, and the score should be considered exploratory until externally validatedAbbreviations: SE Standard Error, HR Hazard Ratio, CI Confidence Interval, SII Systemic Immune-Inflammation Index Multivariate Cox regression analysis and derivation of the CLIP-M score Variables excluded from the final model included sex, ECOG-PS, comorbidity, and liver metastasis. With 63 events and 3 variables in the final refitted model, the final-model EPV was 21. A global Wald test did not reject equality of the three retained coefficients (χ²=0.208, df = 2, p = 0.901), supporting the unit-weighting approach. Candidate inflammatory indices were not entered simultaneously; instead, each index was evaluated in a separate core-adjusted model to reduce redundancy and collinearity. Nevertheless, although the final refitted model had a nominal EPV of 21, this does not fully capture the optimism introduced by candidate biomarker screening, cut-off selection, and backward selection; therefore, residual overfitting cannot be excluded, and the score should be considered exploratory until externally validated Abbreviations: SE Standard Error, HR Hazard Ratio, CI Confidence Interval, SII Systemic Immune-Inflammation Index In Kaplan–Meier–based time-dependent ROC analysis at the prespecified 12-month horizon, SII showed modest discrimination (AUC 0.568), and the optimal Youden cut-off was 789 (sensitivity 0.626, specificity 0.613). The corresponding 12-month time-dependent ROC curve and Youden-derived threshold characteristics are provided in Supplementary Figure S2 and Supplementary Table S5. Using this threshold, 38 patients (47.5%) were classified as low SII and 42 (52.5%) as high SII. In a prespecified multivariable Cox model including age ≥ 65 years, ECOG-PS (dichotomized), gender, primary site group (non-head & neck vs. head & neck), liver metastasis, and comorbidity, SII treated as a continuous variable remained independently associated with OS (p = 0.030), and non-head & neck primary site was associated with inferior OS (HR = 2.230, 95% CI 1.219–4.077; p = 0.009). Because SII had a right-skewed distribution (Supplementary Figure S3), a sensitivity analysis using ln-transformed SII was performed within the same covariate set. In this analysis, the association was attenuated and did not reach statistical significance (HR = 1.361, 95% CI 0.918–2.018; p = 0.125). Based on the comparable hazard ratios of these independent predictors (HR range: 1.86–2.21), the CLIP-M score was constructed by assigning 1 point to each risk factor: SII ≥ 789 (1 point)Age ≥ 65 years (1 point)Non–head & neck primary site (1 point) SII ≥ 789 (1 point) SII ≥ 789 (1 point) Age ≥ 65 years (1 point) Age ≥ 65 years (1 point) Non–head & neck primary site (1 point) Non–head & neck primary site (1 point) The total score ranged from 0 to 3. The distribution of patients according to the CLIP-M score was as follows: 16 patients (20.0%) had a score of 0, 15 (18.8%) had a score of 1, 39 (48.8%) had a score of 2, and 10 (12.5%) had a score of 3. Consistent with Kaplan–Meier patterns and pairwise comparisons showing similar outcomes for scores 0 and 1, whereas scores 2 and 3 were associated with poorer survival, patients were stratified into low-risk (scores 0–1) and high-risk (scores 2–3) groups. In Cox regression modeling CLIP-M as a four-level categorical variable (score 0 as reference), score 1 was not associated with OS (HR 1.19, 95% CI 0.47–3.03; p = 0.719), whereas score 2 (HR 4.10, 95% CI 1.89–8.88; p < 0.001) and score 3 (HR 6.34, 95% CI 2.40–16.73; p < 0.001) were significantly associated with worse OS. Overall, the categorical CLIP-M score was significant (global Wald test p < 0.001). Internal validation In bootstrap internal validation (1,000 resamples), the dichotomized CLIP-M model (low vs. high risk) demonstrated moderate discrimination, with an apparent C-index of 0.648, an optimism-corrected C-index of 0.648, and a calibration slope of 1.011. The bias-corrected and accelerated 95% confidence interval for the log-hazard coefficient of the high-risk group (vs. low-risk) ranged from 0.806 to 2.197 and did not include zero. Bootstrap coefficient stability analyses of the final three-component CLIP-M model showed that the direction of effect was preserved in nearly all successful resamples for high SII (99.7%), age ≥ 65 years (99.2%), and non-head & neck primary site (100.0%), although percentile intervals remained relatively wide, consistent with the modest sample size (Supplementary Table S10). By comparison, the dichotomized RMH score (0 vs. 1–3) showed lower discrimination, with an apparent C-index of 0.602 and an optimism-corrected C-index of 0.601, and a less favorable calibration slope (1.145). Bootstrap-corrected calibration plots at 12 months indicated good agreement between predicted and observed survival for CLIP-M, whereas the RMH score tended to overestimate risk in higher-risk patients (Supplementary Figure S1). Validation and comparison with RMH score The prognostic performance of the CLIP-M score was evaluated in a comparative assessment with the RMH score. Distribution of RMH categories was markedly skewed, with 60.0% of patients classified as score 0 and only 13.8% as scores 2–3. Owing to sparse counts in higher categories and to improve clinical interpretability, the RMH score was dichotomized for Kaplan–Meier and Cox group comparisons (score 0 vs. 1–3), whereas ordinal RMH point scores were used for time-dependent AUC analyses. Discriminatory performance was assessed using IPCW time-dependent ROC analyses at prespecified time points and influence-function–based inference. CLIP-M showed numerically higher time-dependent AUCs than RMH at 12 months (AUC 0.731 vs. 0.613), 24 months (0.824 vs. 0.680), and 36 months (0.863 vs. 0.697) (Fig. 2). The between-model difference reached nominal statistical significance only at 36 months (ΔAUC = 0.166; nominal p = 0.039), but this did not remain significant after correction for the three prespecified time horizons. Although the number of patients at risk decreased at later time horizons, these time points were defined a priori to avoid data-driven selection. Fig. 2Time-dependent AUC estimates (with 95% CI) comparing CLIP-M versus RMH at 12, 24, and 36 months for overall survival. Higher AUC indicates better discrimination Time-dependent AUC estimates (with 95% CI) comparing CLIP-M versus RMH at 12, 24, and 36 months for overall survival. Higher AUC indicates better discrimination Clinically, CLIP-M provided clear risk stratification, separating patients into low-risk (0–1 point) and high-risk (2–3 points) groups with markedly different outcomes. Median overall survival was 24.8 months in the low-risk group versus 6.4 months in the high-risk group (log-rank p < 0.001) (Fig. 3A), corresponding to an approximately fourfold increase in mortality risk (HR 4.03; 95% CI, 2.25–7.22). In contrast, the RMH score showed less pronounced separation (median OS, 12.1 vs. 6.1 months; log-rank p = 0.003) (Fig. 3B). Fig. 3Kaplan–Meier OS estimates for patients stratified by the CLIP-M and RMH score. Kaplan–Meier curves were truncated at 48 months because of the limited number of patients remaining at risk beyond this time point. A CLIP-M clearly separated low-risk (score 0–1) and high-risk (score 2–3) patients, with a median OS of 24.8 vs. 6.4 months (log-rank p < 0.001). B The RMH score showed a less pronounced separation between dichotomized risk groups. Median OS was 12.1 versus 6.1 months for the low- and high-risk groups, respectively (log-rank p = 0.003) Kaplan–Meier OS estimates for patients stratified by the CLIP-M and RMH score. Kaplan–Meier curves were truncated at 48 months because of the limited number of patients remaining at risk beyond this time point. A CLIP-M clearly separated low-risk (score 0–1) and high-risk (score 2–3) patients, with a median OS of 24.8 vs. 6.4 months (log-rank p < 0.001). B The RMH score showed a less pronounced separation between dichotomized risk groups. Median OS was 12.1 versus 6.1 months for the low- and high-risk groups, respectively (log-rank p = 0.003) In direct Cox regression comparison, CLIP-M remained strongly and independently associated with overall survival (HR 3.66; 95% CI, 2.03–6.63; p < 0.001), whereas the RMH score demonstrated weaker, albeit statistically significant, prognostic value (HR 1.70; 95% CI, 1.01–2.84; p = 0.044) (Table 3). Table 3Comparative performance of CLIP-M and RMH scoresMetricCLIP-MRMHBetween-model differencesAUC at 12 months (95% CI)0.731 (0.620–0.843)0.613 (0.500–0.727)ΔAUC = 0.118; p = 0.111AUC at 24 months (95% CI)0.824 (0.714–0.934)0.680 (0.573–0.787)ΔAUC = 0.144; p = 0.078AUC at 36 months (95% CI)0.863 (0.754–0.971)0.697 (0.594–0.799)ΔAUC = 0.166; nominal p = 0.039Median OS in low vs. high risk groups24.8 vs. 6.4 months (p < 0.001)12.1 vs. 6.1 months (p = 0.003)—Direct comparative Cox HR (95% CI)3.66 (2.03–6.63)1.70 (1.01–2.84)—Direct comparative Cox p-value< 0.0010.044—Numbers at risk: 12 m = 28, 24 m = 13, 36 m = 10. Time-dependent AUCs were calculated using ordinal point scores, whereas median OS and Cox HRs were based on dichotomized risk groups. Between-model ΔAUC p-values were calculated using influence-function–based inference; given the three prespecified time horizons, the 36-month p-value should be interpreted as nominal and did not remain significant after correction for multiple time-point comparisonsAbbreviations: AUC Area Under the Curve, CI Confidence Interval, CLIP-M Clinical and Inflammatory Prognostic Score in Mucosal Melanoma, HR Hazard ratio, OS Overall Survival, RMH Royal Marsden Hospital Table 3Comparative performance of CLIP-M and RMH scoresMetricCLIP-MRMHBetween-model differencesAUC at 12 months (95% CI)0.731 (0.620–0.843)0.613 (0.500–0.727)ΔAUC = 0.118; p = 0.111AUC at 24 months (95% CI)0.824 (0.714–0.934)0.680 (0.573–0.787)ΔAUC = 0.144; p = 0.078AUC at 36 months (95% CI)0.863 (0.754–0.971)0.697 (0.594–0.799)ΔAUC = 0.166; nominal p = 0.039Median OS in low vs. high risk groups24.8 vs. 6.4 months (p < 0.001)12.1 vs. 6.1 months (p = 0.003)—Direct comparative Cox HR (95% CI)3.66 (2.03–6.63)1.70 (1.01–2.84)—Direct comparative Cox p-value< 0.0010.044—Numbers at risk: 12 m = 28, 24 m = 13, 36 m = 10. Time-dependent AUCs were calculated using ordinal point scores, whereas median OS and Cox HRs were based on dichotomized risk groups. Between-model ΔAUC p-values were calculated using influence-function–based inference; given the three prespecified time horizons, the 36-month p-value should be interpreted as nominal and did not remain significant after correction for multiple time-point comparisonsAbbreviations: AUC Area Under the Curve, CI Confidence Interval, CLIP-M Clinical and Inflammatory Prognostic Score in Mucosal Melanoma, HR Hazard ratio, OS Overall Survival, RMH Royal Marsden Hospital Comparative performance of CLIP-M and RMH scores Numbers at risk: 12 m = 28, 24 m = 13, 36 m = 10. Time-dependent AUCs were calculated using ordinal point scores, whereas median OS and Cox HRs were based on dichotomized risk groups. Between-model ΔAUC p-values were calculated using influence-function–based inference; given the three prespecified time horizons, the 36-month p-value should be interpreted as nominal and did not remain significant after correction for multiple time-point comparisons Abbreviations: AUC Area Under the Curve, CI Confidence Interval, CLIP-M Clinical and Inflammatory Prognostic Score in Mucosal Melanoma, HR Hazard ratio, OS Overall Survival, RMH Royal Marsden Hospital Bootstrap distributions of IPCW time-dependent AUC estimates, including mean, median, and percentile 95% confidence intervals, are provided in Supplementary Table S8. In single-score Cox models, CLIP-M showed better relative model fit than RMH, with a lower AIC (436.028 vs. 452.708) and a larger likelihood-ratio chi-square statistic (24.909 vs. 8.229). These findings were consistent with the numerically higher discrimination observed for CLIP-M in time-dependent AUC analyses (Supplementary Table S3). In exploratory incremental analyses, SII provided additional prognostic information beyond the ordinal RMH score. In the RMH-only Cox model, RMH score was associated with OS (HR 1.676 per 1-point increase, 95% CI 1.225–2.292; p = 0.001; AIC 451.666). After adding dichotomized SII, both RMH score (HR 1.526, 95% CI 1.098–2.121; p = 0.012) and high SII (HR 1.771, 95% CI 1.040–3.017; p = 0.035) remained associated with OS, with improved model fit (AIC 449.177; likelihood-ratio test for added SII, p = 0.034). Apparent time-dependent AUCs also increased after adding SII to RMH at 12 months (0.613 to 0.693), 24 months (0.680 to 0.761), and 36 months (0.697 to 0.827), although these incremental discrimination analyses were exploratory and derived within the same dataset (Supplementary Table S11). For descriptive prediction of 12-month mortality, 72 patients were evaluable after excluding 8 patients censored before 12 months. The dichotomized CLIP-M score showed a sensitivity of 79.5%, specificity of 64.3%, PPV of 77.8%, NPV of 66.7%, and overall accuracy of 73.6% (Supplementary Table S2). Exploratory 12-month decision curve analysis showed that CLIP-M and RMH provided similar net benefit at lower threshold probabilities, whereas CLIP-M showed modestly higher net benefit at intermediate threshold probabilities, particularly around 0.40–0.50 for 12-month mortality. At a threshold probability of 0.50, the net benefit of CLIP-M (0.305) exceeded that of RMH (0.164) and the treat-all strategy (0.152) (Supplementary Figure S5). First-line systemic therapy distribution was examined according to age group and CLIP-M risk group. Treatment modality was not significantly associated with age ≥ 65 years (Fisher–Freeman–Halton exact p = 0.760) or CLIP-M risk group (p = 0.755) (Supplementary Table S6). In a treatment-adjusted Cox model including first-line systemic therapy, CLIP-M high-risk status remained associated with inferior OS (HR 3.885, 95% CI 2.162–6.982; p < 0.001). Compared with chemotherapy, immunotherapy was not significantly associated with OS (HR 1.227, 95% CI 0.678–2.223; p = 0.499), whereas no systemic therapy/best supportive care was associated with inferior OS (HR 3.257, 95% CI 1.354–7.839; p = 0.008) (Supplementary Table S7). Among patients receiving active systemic therapy, no statistically significant interaction was observed between CLIP-M risk group and treatment modality (chemotherapy vs. immune checkpoint inhibitors; interaction p = 0.896). However, this interaction analysis was underpowered because only 20 patients received immunotherapy. Calendar-era sensitivity analyses were performed using the date of advanced disease diagnosis. First-line systemic therapy distribution differed significantly across calendar eras (Fisher–Freeman–Halton exact p = 0.023), with immunotherapy use increasing from 3/32 patients (9.4%) in 2011–2017 to 17/48 patients (35.4%) in 2018–2025. CLIP-M risk group distribution did not significantly differ by era (Fisher exact p = 0.350). In an era-adjusted Cox model, CLIP-M high-risk status remained associated with inferior OS (HR 3.760, 95% CI 2.082–6.791; p < 0.001), whereas calendar era was not independently associated with OS (HR 0.761, 95% CI 0.456–1.268; p = 0.294). No statistically significant CLIP-M × calendar-era interaction was observed (interaction HR 1.957, 95% CI 0.642–5.968; p = 0.238). In era-stratified analyses, CLIP-M remained prognostic in both 2011–2017 (HR 2.497, 95% CI 1.085–5.746; p = 0.031) and 2018–2025 (HR 8.620, 95% CI 3.115–23.858; p < 0.001), although subgroup estimates should be interpreted cautiously (Supplementary Table S9). Discussion Advanced MMM is clinically aggressive and associated with poor survival, underscoring the need for clinically interpretable risk stratification tools. In this study, we derived and internally validated a novel clinicobiological prognostic model, the CLIP-M score, specifically tailored for patients with advanced MMM. The CLIP-M score—comprising SII, age, and tumor location—showed numerically higher apparent discrimination than the RMH score, although optimism-corrected performance remained moderate. To our knowledge, this is among the first studies to construct an internally validated composite scoring system for advanced MMM by integrating systemic inflammatory burden with clinical and anatomical variables. The recognized shortcomings of conventional staging systems in MMM have prompted increasing interest in biomarkers that better capture host–tumor biological interactions. Although our group has previously demonstrated the prognostic relevance of SII as a standalone biomarker in a broader MMM population [11], the current study specifically addresses the advanced-disease setting and extends the prior work by evaluating candidate inflammatory and nutritional indices for integration with clinical and anatomical variables into an internally validated composite score. In the candidate-marker analyses, both SII and GPS showed prognostic relevance. However, GPS is itself an established albumin/CRP-based prognostic score and partially overlaps conceptually with albumin-containing scores such as RMH [20, 23]. Therefore, we did not incorporate GPS as a component of the new score. Instead, SII was selected as the biological component of CLIP-M because it captures a distinct hematologic immune-inflammatory axis based on platelet, neutrophil, and lymphocyte counts, remained independently associated with OS when modeled continuously, and could be translated into a simple binary component using a prespecified 12-month time-dependent ROC threshold. The attenuation of the association after ln-transformation suggests that the prognostic signal of SII may be influenced by higher-range values; therefore, SII should be interpreted as one component of a composite prognostic score rather than as a standalone continuous biomarker. These findings also underscore the conceptual limitation of relying on a single baseline SII cut-off in a biologically dynamic disease setting. Emerging melanoma data suggest that on-treatment inflammatory kinetics, such as early CRP flare after immune checkpoint inhibitor initiation, may capture treatment-related host–tumor immune dynamics; therefore, future studies should evaluate whether serial SII or CRP kinetics add prognostic or predictive value beyond a baseline score [24]. For bedside applicability, the SII threshold was derived within the development cohort. Notably, a similar SII cut-off has been reported in an independent immunotherapy-treated solid-tumor cohort [25]; however, because that cohort involved a different disease context, this observation should be viewed only as contextual support rather than external validation. This finding aligns with recent literature suggesting that in highly immunogenic or aggressive tumors like melanoma, the intensity of the systemic inflammatory response and immune evasion mechanisms may be more dominant drivers of mortality than nutritional status alone [26]. Like the LIPI in non-small cell lung cancer, CLIP-M highlights the critical role of systemic inflammation in survival outcomes [19]. A key novelty of the CLIP-M score is the integration of age and tumor location as independent risk factors. Interestingly, while age is often not a significant factor in localized MMM, our results indicate that in the advanced setting, patients aged ≥ 65 years have a significantly worse prognosis. This likely reflects reduced physiological reserve (frailty) and immunosenescence, which may compromise the host’s ability to mount an effective antitumor response or tolerate systemic therapies in the face of high tumor burden [27]. Similarly, the distinction between head & neck and non–head & neck primary sites was clinically relevant. Consistent with previous studies suggesting distinct biological behaviors and metastatic patterns [14, 15], non–head & neck primaries were associated with approximately a two-fold increase in mortality risk, supporting their inclusion in the high-risk category of the model. The comparison with the RMH score provides a clinically relevant benchmark, although the RMH score was originally developed for patients enrolled in phase I trials across heterogeneous advanced solid tumors rather than specifically for MMM [7]. Although RMH and related prognostic tools have been evaluated in advanced melanoma populations, they were not developed to capture the distinct biology of mucosal melanoma [28, 29]. The RMH score incorporates LDH, an established adverse prognostic biomarker across solid tumors, but does not directly capture host inflammatory status [30]. In exploratory incremental analyses, SII remained associated with OS after adjustment for the ordinal RMH score and modestly improved model fit and apparent time-dependent discrimination. This supports the concept that systemic inflammatory status may provide prognostic information not fully captured by RMH components such as LDH, albumin, and metastatic-site burden. In our cohort, time-dependent AUC estimates numerically favored CLIP-M over RMH at 12, 24, and 36 months, and CLIP-M also showed clearer separation of low- and high-risk groups in Kaplan–Meier analyses. Consistent with these findings, exploratory 12-month decision curve analysis suggested modestly greater clinical net benefit for CLIP-M than for RMH at intermediate threshold probabilities, particularly around 0.40–0.50. However, these findings should be interpreted cautiously. The time-dependent AUC estimates reflect the ability of each score to discriminate between patients who died before a given time horizon and those who survived beyond that horizon while accounting for censoring. Later time-point estimates, particularly at 36 months, are less stable because of the small number of patients remaining at risk. Therefore, the 36-month AUC difference should be considered exploratory rather than definitive evidence of long-term superiority. The higher apparent time-dependent AUCs should also be interpreted in light of the more modest optimism-corrected C-index, as these metrics capture different aspects of discrimination and are affected differently by sample size, censoring, and internal resampling. Accordingly, the apparent AUC advantage of CLIP-M should be considered hypothesis-generating until externally validated. CLIP-M should not be interpreted as a treatment-selection tool in its current form. In the absence of external validation and predictive interaction data, the score should not be used to choose single-agent versus combination immunotherapy. Its potential clinical role is more appropriately framed as baseline prognostic stratification: identifying patients who may require closer monitoring, earlier response assessment, supportive-care optimization, or prioritization for clinical trials. Whether CLIP-M can guide treatment intensification requires prospective validation in contemporary immunotherapy-treated cohorts. This study has several limitations, primarily related to its retrospective design and the rarity of the disease. A complete-case approach was used; 31 of 111 (27.9%) otherwise eligible patients were excluded because baseline clinical and/or laboratory data required for index calculation and candidate predictors were missing. Because these excluded patients were not systematically abstracted into the analytic dataset beyond eligibility screening, an included-versus-excluded baseline comparison could not be reliably performed. This may introduce selection bias and limit generalizability. Multiple imputation was not used because missingness involved source laboratory variables required to construct multiple derived inflammatory and nutritional indices, and imputation in this small cohort could have introduced additional uncertainty into score development. Comorbidity was captured as a binary variable, and detailed comorbidity subtypes were not available; therefore, the potential influence of specific comorbid conditions on inflammatory indices could not be examined. Molecular testing was incomplete and heterogeneous across centers and calendar periods, precluding reliable incorporation of mutation status into model development or sensitivity analyses. The non–head & neck category combined gastrointestinal, gynecological, urinary/genitourinary, and rare respiratory mucosal primaries; however, site-level subgroup analyses were not feasible because of the small number of patients in each non–head & neck subsite. Although time-dependent AUC was used to account for censoring, the sample size remained modest, resulting in relatively wide confidence intervals. The broader model-development process also involved multiple candidate assessments, cut-off selection, and backward selection. Therefore, residual optimism remains possible despite the acceptable nominal EPV of the final refitted model. The incremental RMH-plus-SII analyses were also exploratory apparent-performance analyses performed within the same dataset; therefore, they require external validation. The decision curve analysis was also exploratory and performed in the derivation cohort; therefore, the apparent net-benefit advantage of CLIP-M should be interpreted cautiously pending external validation. Treatment heterogeneity over the long inclusion period may also confound prognostic estimates. Treatment-adjusted and interaction analyses were performed, but these analyses were limited by the small number of immunotherapy-treated patients and should be considered exploratory. Calendar-era analyses were also exploratory; although CLIP-M remained prognostic after era adjustment and in era-stratified models, these findings should not be interpreted as definitive evidence of temporal stability because subgroup sizes were limited and residual confounding by evolving systemic therapy, supportive care, and treatment access remains possible. In addition, palliative radiotherapy exposure was not systematically recorded, and radiotherapy-based treatments were heterogeneous with respect to timing and indication; therefore, their potential survival or immunomodulatory impact could not be fully modeled. The nominal 36-month AUC difference should be interpreted with caution because only 10 patients remained at risk at this time point and the p-value did not remain significant after correction for multiple time-point comparisons. Consequently, long-term discrimination estimates may be unstable and should be considered exploratory. External validation in independent and preferably prospective cohorts is required before widespread clinical adoption. Conclusion In patients with advanced MMM, the CLIP-M score provided clinically interpretable prognostic stratification and remained associated with OS after internal validation. However, given the modest sample size, complete-case design, treatment heterogeneity, and lack of external validation, CLIP-M should be regarded as an exploratory prognostic tool. Independent validation in contemporary cohorts is required before clinical implementation. Supplementary Information Supplementary Material 1. Supplementary Material 1.