Abstract
BACKGROUND: Effective management of borderline ovarian tumors (BOTs) requires timely identification of patients at high risk of recurrence. Previous studies suggest that artificial neural networks can improve the prediction of BOT recurrence compared to traditional models, though concerns about their validity persist due to insufficient external validation. We aimed to evaluate the predictive performance of a time-dependent artificial neural network, conduct comprehensive temporal and spatial external validations to address this critical limitation. METHODS: Clinical data were collected from patients diagnosed with BOT at Shengjing Hospital of China Medical University between January 2014 and August 2023, including 76 cases of recurrence and 584 cases of non-recurrence. Using the Synthetic Minority Oversampling Technique (SMOTE), we balanced the groups at a 1:1 ratio (total sample size, N = 1168). Random sampling was used to divide the data into a training set (70%) and an internal validation set (30%). Temporal (same center; May 2011-December 2013) and spatial (different center; April 2011-April 2019) external validation sets were established. The training set data were input into the input layer of the neural network, high-level features were extracted via neurons in the hidden layer, and the model output was generated from the output layer. The model's sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, time-dependent area under the receiver operating characteristic curve (tdAUC), and integrated Brier score (IBS) were evaluated using the internal validation set, the temporal external validation set, and the spatial external validation set, respectively. RESULTS: A neural multi-task logistic regression model (N-MTLR) was constructed based on 34 features from the training set after correlation screening, with 9 variables selected for the neural network prediction model. The prediction model consisted of three functional layers with 128, 64, and 32 neurons, respectively, totaling 224 neurons. The optimal parameters for the final model were set as follows: initialization method of glorot_uniform, Dropout rate of 30%, L2 regularization parameter of 1e-2, optimizer of Adam, and learning rate of 1e-4. The N-MTLR model produced higher predictive performance including AUC, accuracy, specificity, PPV and NPV than the Cox-regression model for all survival endpoints at the 2-, the 4- and 7-year time points. Both temporal and spatial external validation results indicated that the model had moderate predictive performance and certain clinical application value. CONCLUSIONS: The N-MTLR neural network enables superior nonlinear modeling of BOT recurrence risk, exhibits excellent temporal and spatial generalizability, which supports precise risk stratification for clinical decision-making.