An Innovative Approach for Predicting Prostate Cancer Gleason Grading: Machine Learning-based Fusion of Multimodal Ultrasound, Clinical and Laboratory Indicators

Background: Prostate cancer is a common malignancy among elderly males with a growing incidence. While prostate biopsy remains the gold standard for diagnosis, this invasive procedure is poorly tolerated by some patients. The Gleason grade group (GGG) plays a critical role in predicting metastatic risk, guiding treatment selection, and is strongly associated with survival outcomes. Consequently, noninvasive prediction of prostate cancer Gleason grading has emerged as a research priority. This study aimed to develop a noninvasive predictive model integrating multimodal ultrasound data and clinical laboratory biomarkers to preoperatively determine GGGs in prostate cancer patients. Methods: This single-center prospective study enrolled 329 prostate cancer patients meeting predefined inclusion criteria. All participants underwent prostate biopsy with subsequent Gleason grading and were categorized into three groups: low-grade (Gleason score ≤6), intermediate-grade (Gleason score 7), and high-grade (Gleason score ≥8). Thirty-seven predictive parameters were collected, including clinical laboratory biomarkers, systemic inflammatory markers (e.g., neutrophil-to-lymphocyte ratio), and multimodal ultrasound data: Grayscale sonographic characteristics, contrast-enhanced ultrasound (CEUS) parameters, elastography parameters, and radiofrequency signal data. Following feature selection, five clinically significant predictors were identified. Multiple machine learning algorithms were implemented for predictive modeling, and model performance was quantified using accuracy, recall, and F1-score. Results: Six machine learning-based predictive models were developed and evaluated. The Decision Tree model achieved an accuracy of 0.818, recall of 0.818, and F1-score of 0.816. The Random Forest classifier demonstrated an accuracy of 0.820, recall of 0.820, and F1-score of 0.820. The K-Nearest Neighbors algorithm yielded an accuracy of 0.788, recall of 0.788, and F1-score of 0.801. The Gradient Boosting Decision Tree (GBDT) model exhibited superior predictive capability with an accuracy of 0.848, recall of 0.848, and F1-score of 0.849. The XGBoost algorithm had an accuracy of 0.818, recall of 0.789, and F1-score of 0.796, while the Naive Bayes classifier attained an accuracy of 0.773, recall of 0.773, and F1-score of 0.779. Comparative analysis revealed that the GBDT model demonstrated optimal performance among the evaluated algorithms, suggesting its potential clinical significance in predicting Gleason grades. Conclusion : Ultrasonography, being noninvasive, radiation-free, and cost-effective, demonstrates high clinical feasibility for implementation in routine practice, particularly in primary healthcare settings. The predictive model established through multimodal ultrasound parameters effectively predicts the Gleason grade of prostate cancer.