Neural Network Accurately Predicts Lymph Node Metastasis Before Prostate Cancer Surgery

An artificial neural network demonstrated high diagnostic accuracy in preoperatively predicting lymph node metastasis among prostate cancer patients undergoing radical prostatectomy with extended pelvic lymph node dissection, achieving an accuracy of 94.10 percent and an area under the receiver operating characteristic curve of 0.962 in an independent test dataset. Conducted at a tertiary referral center and published in the African Journal of Urology, the retrospective study evaluated whether nonlinear machine learning models could improve individualized risk stratification compared to established nomograms. In prostate cancer management, detecting lymph node metastasis is a vital prognostic determinant that directly influences postoperative treatment decisions and long-term oncological outcomes. Although extended pelvic lymph node dissection represents the most reliable staging method during radical prostatectomy, its routine application is constrained by increased operative complexity and perioperative surgical complications. Consequently, selective implementation guided by individualized risk estimation is essential for optimal surgical planning.

Currently, clinical risk estimation for lymph node metastasis relies on validated statistical nomograms such as the Briganti, Partin, and Memorial Sloan Kettering Cancer Center instruments. While these clinical prediction tools are widely integrated into surgical practice, they are founded upon predefined linear assumptions regarding relationships among predictor variables. In reality, prostate cancer progression is biologically heterogeneous and involves complex, nonlinear interactions between primary tumor volume, histopathological aggressiveness, and radiographic imaging findings. Standard regression models may have limited capacity to capture these multidimensional biological dynamics, potentially curtailing predictive performance across diverse patient groups. In contrast, machine learning encompasses data-driven modeling approaches that function without prespecified assumptions regarding variable relationships. By flexibly modeling nonlinear interactions among clinical parameters, machine learning techniques may provide distinct advantages over conventional regression-based tools. Although artificial intelligence applications have expanded across urologic oncology, published evidence specifically evaluating machine learning for preoperative nodal metastasis prediction in radical prostatectomy patients has remained comparatively scarce.

To address this gap, the researchers conducted a retrospective analysis of consecutive prostate cancer patients who underwent radical prostatectomy with concomitant extended pelvic lymph node dissection between January 2017 and January 2021 at a single tertiary referral center. The institutional ethics committee approved the study protocol under approval number 3367. To be eligible, patients were required to have histopathologically confirmed prostate cancer as well as comprehensive preoperative clinical, biopsy, and multiparametric magnetic resonance imaging records. The study protocol applied strict exclusion criteria to avoid therapeutic confounding. Patients who had undergone prior pelvic radiotherapy, systemic chemotherapy, or androgen deprivation therapy were excluded from participation. Similarly, individuals with documented distant metastasis, active systemic infection, concurrent autoimmune disease, chronic corticosteroid therapy, or incomplete medical records were omitted. Following the application of these criteria, the final study population comprised 110 consecutive patients whose complete medical profiles permitted comprehensive predictive modeling and subsequent histopathological verification.

The investigators collected baseline demographic, clinical, laboratory, and radiographic data for all 110 patients in the cohort. The mean age was 62.14 ± 6.43 years, with a mean preoperative serum prostate-specific antigen level of 10.84 ± 11.39 nanograms per milliliter. The cohort exhibited a mean prostate-specific antigen density of 0.22 ± 0.22 nanograms per milliliter per cubic centimeter and a mean prostate volume of 57.08 ± 26.53 cubic centimeters. On preoperative multiparametric magnetic resonance imaging, 63 patients, representing 57.27 percent of the cohort, had Prostate Imaging Reporting and Data System version 2 scores of four or higher. Biopsy evaluations showed that 58 patients, or 52.73 percent, had an International Society of Urological Pathology grade group of two or greater, while the mean percentage of positive biopsy cores was 37.89 ± 19.57 percent. Baseline preoperative risk estimation using the Briganti nomogram yielded a mean predicted lymph node metastasis probability of 5.73 ± 7.05 percent across the study cohort.

Postoperative histopathological analysis of surgical specimens established the definitive pathological characteristics of the study population. Extraprostatic extension was observed in 35 patients, representing 31.82 percent of the cohort, while seminal vesicle invasion occurred in 12 patients, or 10.91 percent. Perineural invasion was documented in 34 patients, corresponding to 30.91 percent, and pathological stage pT3 or higher was confirmed in 43 patients, representing 39.09 percent of cases. The primary study endpoint was histopathologically proven lymph node metastasis in tissue obtained during extended pelvic lymph node dissection. Pathological examination confirmed nodal metastasis in seven patients, establishing a nodal involvement rate of 6.36 percent in this surgically treated series. These seven positive cases formed the confirmed event group for model development and evaluation. Because all relevant parameters were fully documented across the cohort, no missing values were present in the dataset, and no imputation techniques were required during data preparation.

Predictive model development and reporting complied with the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis–Artificial Intelligence guidelines, utilizing MATLAB R2024b and IBM SPSS Statistics version 21. Continuous predictor variables were normalized prior to model training. To identify the most informative predictors and minimize overfitting risk, feature importance was evaluated using three independent ranking methods: Chi-square test, analysis of variance, and the Kruskal–Wallis test. All three ranking methods demonstrated consistent results, identifying tumor burden and histopathological grade parameters as the most influential predictors of lymph node metastasis. The number of positive biopsy cores exhibited the highest importance scores across algorithms, with an analysis of variance score of 22.31 and a Kruskal–Wallis score of 16.52. The International Society of Urological Pathology grade group followed, achieving an analysis of variance score of 19.65 and a Kruskal–Wallis score of 15.32, while the percentage of positive biopsy cores yielded scores of 15.18 and 12.80, respectively. Clinical tumor stage and perineural invasion demonstrated moderate predictive importance, whereas imaging parameters, age, and prostate-specific antigen density scored lower. Body mass index consistently ranked as the least informative predictor.

The researchers evaluated several machine learning architectures, including logistic regression, decision trees, support vector machines, naïve Bayes, boosted trees, and artificial neural networks. For the artificial neural network, the authors implemented a feedforward architecture comprising one hidden layer configured with a sigmoid activation function and an output layer utilizing a softmax function. Network parameters and neuron counts were empirically optimized against validation performance, using backpropagation with a gradient-based optimization algorithm. Early stopping based on validation loss was incorporated to curtail overfitting, and no data augmentation techniques were employed. The patient cohort was randomly divided into a training subset of 76 patients, a validation subset of 17 patients, and an independent test subset of 17 patients for neural network evaluation. For the remaining machine learning algorithms, model validation was executed using five-fold cross-validation. Diagnostic performance across all models was assessed through confusion matrices and receiver operating characteristic curve analysis, evaluating accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and area under the curve.

Among the evaluated techniques, the artificial neural network achieved the highest diagnostic performance for predicting lymph node metastasis. In the independent test dataset of 17 patients, the artificial neural network attained an accuracy of 94.10 percent and an area under the curve of 0.962. In this test cohort, the model demonstrated a sensitivity of 1.000, a specificity of 0.923, a positive predictive value of 0.800, and a negative predictive value of 1.000. Within the training dataset of 76 patients, the neural network achieved an area under the curve of 0.967, a sensitivity of 0.950, a specificity of 0.946, a positive predictive value of 0.864, and a negative predictive value of 0.981. In the validation dataset of 17 patients, the model recorded an area under the curve of 0.846, with sensitivity and specificity reaching 0.750 and 0.769, alongside a positive predictive value of 0.500 and a negative predictive value of 0.909. Across all datasets combined, the overall area under the curve was 0.955, with an overall sensitivity of 0.929, an overall specificity of 0.915, a positive predictive value of 0.788, and a negative predictive value of 0.974, confirming strong internal discrimination.

When compared with alternative machine learning techniques, the artificial neural network consistently exhibited superior discrimination. Support vector machines achieved an area under the curve of 0.90, while multivariable logistic regression attained an area under the curve of 0.88. Naïve Bayes generated the highest validation accuracy among alternative models at 86.17 percent. Decision tree and boosted tree models displayed lower overall performance, recording accuracies of approximately 79 percent, with area under the curve values ranging from 0.74 to 0.87. Furthermore, the neural network demonstrated superior discrimination when compared with risk stratification derived from the Briganti nomogram. The authors highlighted that the high negative predictive value of the neural network—reaching 1.000 in testing and 0.974 overall—represents a finding of meaningful clinical relevance. Accurately identifying patients who possess a minimal probability of nodal involvement could allow urologists to selectively omit extended pelvic lymph node dissection, thereby reducing surgical morbidity while preserving staging accuracy. While these findings align with prior comparative literature indicating that machine learning can match or exceed traditional nomograms, the authors noted that machine learning should serve as a complementary decision-support tool rather than an unverified replacement.

The authors recognized several important limitations that temper these conclusions and underscore the preliminary nature of the findings. The retrospective single-center design introduces inherent risks of selection bias and precludes causal inferences. Furthermore, the study cohort comprised 110 patients with only seven histopathologically confirmed lymph node metastasis events, representing an event rate of 6.36 percent. This small number of positive outcomes may influence model stability, calibration accuracy, and the generalizability of performance estimates despite internal validation safeguards. In addition, the black-box nature of artificial neural networks limits interpretability compared with conventional statistical regression models, which could pose obstacles for clinical acceptance and regulatory adoption. The study was also limited to predicting anatomical lymph node metastasis and did not evaluate downstream oncological or functional outcomes, leaving its real-world impact on clinical decision-making unproven. Future investigations must conduct external validation across larger multicenter cohorts, perform decision-curve analyses, and evaluate whether integrating radiomics, genomic classifiers, and molecular biomarkers can enhance predictive robustness.

Subject of Research: Preoperative artificial neural network prediction of prostate cancer lymph node metastasis

Article Title: Artificial neural network–based prediction of lymph node metastasis in prostate cancer patients undergoing radical prostatectomy

Article References: Albayrak, A. T., Gunay, K. C., Baloglu, I. H., Kutsal, C., Yavuzsan, A. H., Kirecci, S. L., Guney, S., & Dogu, E. (2026). Artificial neural network–based prediction of lymph node metastasis in prostate cancer patients undergoing radical prostatectomy. African Journal of Urology, 32(1), Article 30. https://doi.org/10.1186/s12301-026-00583-5

Image Credits: AI Generated

DOI: 10.1186/s12301-026-00583-5

Keywords: prostate cancer, radical prostatectomy, lymph node metastasis, extended pelvic lymph node dissection, artificial neural networks, machine learning, Briganti nomogram, multiparametric MRI, PI-RADS, prostate-specific antigen, ISUP grade group, biopsy core involvement, predictive modeling, area under the ROC curve, negative predictive value, feature importance ranking, urologic oncology, TRIPOD-AI guidelines