Construction of an oligometastatic prediction model for nasopharyngeal carcinoma patients based on pathomics features and dynamic multi-swarm particle swarm optimization support vector machine.

Objective: This study aimed to develop a risk prediction model for post-treatment oligometastasis in nasopharyngeal carcinoma (NPC) by integrating pathomics features and an improved Support vector machine (SVM) algorithm, offering precise early decision support. Methods: This study retrospectively included 462 NPC patients, without or with oligometastasis defined by ESTRO/EORTC criteria. Whole-slide images were scanned, and three representative H&E-stained regions were selected for pathomics feature extraction via CellProfiler software. Features screened by intraclass correlation coefficient, Mann-Whitney U test, Spearman correlation, minimum redundancy maximum relevance, and Least absolute shrinkage and selection operator regression. Based on these screened features, three models were built: Dynamic Multi-Swarm Particle Swarm Optimization SVM (DMS-PSO-SVM), Particle Swarm Optimization SVM (PSO-SVM), and a standard SVM. Model training and hyperparameter tuning were conducted on the training set (n=369), followed by evaluation on a validation set (n=93). Results: 6 pathomics features were screened as important features. DMS-PSO-SVM yielded superior performance, with training-set AUC=0.880 and validation-set AUC=0.866, consistently outperforming both PSO-SVM (AUC=0.721) and standard SVM (AUC=0.718). Calibration curves showed good agreement for DMS-PSO-SVM (P>0.05) but indicated miscalibration in the standard SVM (P<0.05). Decision curve analysis further demonstrated that DMS-PSO-SVM offered higher net benefit across a wide range of risk thresholds. Conclusion: Incorporating pathomics and DMS-PSO optimization significantly improved NPC oligometastasis prediction. This model showed high discriminative ability, calibration, and clinical utility, suggesting that pathomics and machine learning-based strategies could aid early recognition of high-risk patients and inform individualized treatment approaches. A demo of the DMS-PSO-SVM modeling algorithm code used in this study can be found on Github (https://github.com/Edward-E-S-Wang/DMS-PSO-SVM).