Preliminary modeling of brain metastases in breast cancer: a neural network approach to risk factor analysis

Breast cancer accounts for one third of cancer-related deaths in women and has complications such as bone and brain metastases in other organs. Investigating the factors influencing this disease and its associated metastases is particularly important for treatment trends. Therefore, this study aimed to investigate the predictive risk factors and their risk scores for brain metastases in breast cancer patients. This case-control study was conducted on breast cancer patients in two groups: those with (13 subjects) and those without (39 subjects) metastases, in two medical centres between the 2022–2024. The data of patients were obtained from their medical records. Chi-square tests, multivariable logistic regression, Receiver Operating Characteristic (ROC) curves, and a two-layer perceptron neural network model (NNM) with a hyperbolic tangent activation function were employed to compare differences and predict risk factors of brain metastasis. All data were analysed using SPSS v.28 software. In univariable logistic regression analyses, among significant predictors i.e. younger age (≤ 40 years), HER2–HR+ status, tumor size, and lymph node involvement which inclufed in multivariable analysis younger age and HER2–HR+ status remained independently significant predictors of brain metastasis (BBM). A point-based scoring system with a cutoff of 8 achieved sensitivity of 77%, specificity of 62%, and with an overall discriminative ability of AUC = 0.88 (95% CI: 0.78–0.98, P < 0.001). Neural network analysis provided complementary insights by quantifying the relative importance of predictors. HER2-HR+ (100% relative importance) and progesterone receptor status (79% relative importance) emerged as the most influential variables. The model achieved an AUC of 0.843, with test accuracy of 91.7%. Internal validation confirmed moderate but consistent performance but some miscalibration due to sample size limitations. The scoring model developed in this study demonstrated acceptable sensitivity and specificity for predicting the occurrence of brain metastases in women with breast cancer. However, given the hospital-based design and relatively small sample size, the estimates may be unstable and should be interpreted with caution. These findings suggest that the proposed scoring system may serve as a preliminary tool for risk stratification, while highlighting the need for further refinement and validation in diverse patient populations.