Clinical Risk Models, Inflammatory–Nutritional Indices and a Neural-Network Survival Model in Metastatic Renal Cell Carcinoma: A Single-Center Retrospective Study

Background/Objectives: In metastatic renal cell carcinoma (mRCC), the International Metastatic Renal Cell Carcinoma Database Consortium (IMDC) and Memorial Sloan Kettering Cancer Center (MSKCC) risk models remain the primary tools for prognostic stratification but may not fully capture outcome heterogeneity in the contemporary treatment era. Systemic inflammatory–nutritional biomarkers such as the pan-immune-inflammation value (PIV) and the hemoglobin–albumin–lymphocyte–platelet (HALP) score have emerged as complementary, readily available prognostic indicators, yet their added value alongside established risk models, nephrectomy status, and modern predictive approaches such as machine learning remains poorly defined in real-world cohorts. This study jointly evaluated established risk models, inflammatory–nutritional biomarkers, and an exploratory neural-network survival model for prognostication in mRCC. Methods: In this retrospective, single-center cohort study, 96 patients with mRCC diagnosed and treated between January 2010 and January 2025 were evaluated, with clinical outcomes retrospectively reviewed up to a data cut-off of 17 November 2025. Clinicopathological characteristics, IMDC and MSKCC risk classification, nephrectomy status, and baseline PIV and HALP values were assessed for their association with progression-free survival (PFS) and overall survival (OS) using Kaplan–Meier and multivariable Cox regression analyses. As an exploratory analysis, a dual-head neural network predicting both the risk of death and survival time was developed; feature screening and hyperparameter tuning were nested within stratified 5-fold cross-validation repeated 10 times, and the network was benchmarked on identical splits against IMDC, MSKCC, Cox, elastic-net Cox and random survival forest models. Results: At a median follow-up of 82.6 months, 77 PFS events and 57 deaths had occurred; median PFS and OS were 17.8 and 46.3 months, respectively. IMDC risk classification was significantly associated with both PFS and OS, whereas MSKCC risk classification was associated with OS but not PFS; PIV and HALP were associated with OS only. In multivariable analyses, nephrectomy status remained independently associated with both PFS and OS, whereas IMDC and MSKCC classifications remained associated with OS in separate adjusted models, although the MSKCC estimates should be interpreted cautiously because of the relatively small poor-risk subgroup. The neural network achieved a nested cross-validated C-index of 0.685 (95% CI, 0.608–0.764), comparable to a Cox model with IMDC risk group and nephrectomy status (0.692) and a random survival forest (0.695), numerically higher than the IMDC risk group alone (0.629) and higher than the MSKCC risk group alone (0.594). Its integrated Brier score was 0.189, compared with 0.164 for the IMDC-plus-nephrectomy Cox model, 0.183 for the random survival forest, and 0.221 for the Kaplan–Meier reference; its calibration slope was 0.54, indicating overly extreme predictions. Conclusions: Established risk models remained informative, particularly for OS, while nephrectomy status was the most consistent prognostic factor across conventional and neural-network analyses; PIV and HALP were associated with overall survival in unadjusted analyses but did not show independent prognostic value. An exploratory neural-network model showed discrimination comparable to conventional Cox-based models; however, its individualized survival-probability estimates had greater prediction error than those of the IMDC-plus-nephrectomy Cox model and were poorly calibrated in slope. Recalibration and external validation in larger, contemporary cohorts are required before clinical implementation.