Combination of Naples prognostic score and pathological factors for evaluating the long-term prognosis of patients with late-onset colorectal cancer: a multicenter machine learning study.

Background: Given the increasingly aging population, precise prognostic models are essential for optimizing postoperative care in patients with late-onset colorectal cancer (LOCRC). Inflammatory, immune, and nutritional factors, along with the Naples prognostic score (NPS), platelet-to-lymphocyte ratio (PLR), and clinicopathological variables, are emerging as predictors of postoperative outcomes in patients with LOCRC. In order to overcome traditional statistical limitations in prognosis, this study aimed to evaluate NPS, PLR, and clinicopathological variables as prognostic factors for LOCRC and incorporate them into a machine learning (ML) model for LOCRC prognosis. Methods: The data from 1,588 patients with postoperative LOCRC and aged ≥50 years were retrospectively collected. Patients were randomly divided into a training set (n=1,090) and an internal validation set (n=468) at a 7:3 ratio, with additional patients (n=420) included for external validation. The NPS score was assessed via SPSS (IBM Corp.) and X-tile software to establish optimal cutoffs and divide patients into three groups: 0, 1, and 2. The Kaplan-Meier method and Cox proportional hazard regression analysis were used to evaluate overall survival (OS). Multivariable Cox regression stepwise identified the independent prognostic factors for patients with LOCRC. ML algorithms, random survival forest (RSF), least absolute shrinkage and selection operator (LASSO), stepwise Cox regression, generalized boosted regression modeling, supervised principal components, survival support vector machine, partial least squares regression for Cox, and Cox-Boost were used to identify predictive models for LOCRC. Model performance was assessed by the concordance index (C-index), area under the curve (AUC), and decision curve analysis (DCA). Results: In the training set, 1,558 patients with LOCRC were included. PLR was strongly associated with survival outcome [hazard ratio (HR) =1.358; P=0.02] and was considered independent prognostic risk factors for LOCRC. NPS group, Union for International Cancer Control (UICC) stage, alpha-fetoprotein level, carcinoembryonic antigen level, carbohydrate antigen 19-9 level, tumor differentiation, pT stage, pN stage, and perineural invasion were associated with survival outcomes (HR =0.390, 0.269, 0.544, 0.728, 0.632, 0.452, 0.109, 0.602, and 0.674, respectively; P<0.05) and were considered independent prognostic protective factors for LOCRC. An ML-based nomogram incorporating NPS, PLR, and clinicopathological variables was developed for the prediction of OS and was validated in the training, internal validation, and external validation sets; the C-index values for the LASSO + RSF model in these sets were 0.872, 0.768, and 0.737, respectively, indicating its superior predictive performance compared to other models. Conclusions: The ML-based nomogram, incorporating NPS, PLR, and clinicopathological factors, effectively predicts long-term OS in patients with LOCRC and can provide a cost-efficient prognostic tool for clinical practice.