Background: Brain metastases, secondary brain tumors originating from cancers elsewhere in the body, can present with focal neurologic deficits without systemic symptoms and are usually evaluated using magnetic resonance imaging (MRI). Multimodal imaging methods are usually used to determine the primary cancer source. Radiomic analysis involves extracting quantitative features from medical images using data-characterization algorithms. Recent advances in deep learning and radiomics, which quantitatively analyze imaging data, offer the potential for improved prediction of primary tumor origin through MR imaging analysis. Using longitudinal MRI data, we developed and validated a mixed-effects neural network model for predicting the primary origin of brain metastases.

Methods: We used the open-source dataset from 5 different institutions made available by Ocaña-Tienda et al. (2023) to build the model. We extracted 2446 radiomic features, out of which 60 features were retained after feature selection using wrapper- and filter-based techniques. For the nested mixed-effects neural network model, we incorporated a dual-layer structure with ReLU activation for the fixed effects and unique embeddings for each cluster, leading to a nested structure. The model was compiled using an Adam optimizer and trained across several epochs with a steadily declining learning rate facilitated by exponential decay. Regularization was implicitly managed through embedding and learning rate adjustments. Model performance was evaluated using receiver operating characteristic curve (ROC) analysis.

Results: Our mixed-effects model demonstrated high accuracy (90.5%), precision (92.7%), recall (87.8%), and F1 score (90.2%), with an area under the ROC (AUC-ROC) curve of 0.944 in identifying the primary tumor site. The nested mixed-effects model was superior to the fixed-effects model with an AUC-ROC of 0.719.

Conclusion: Our mixed-effects neural network model accurately predicts the primary origin of brain metastases with high internal and external validity. Using advanced radiomic feature selection and deep learning techniques. Our model achieved high-performance metrics, thus providing a promising tool for enhancing diagnostic capability and tailoring patient-specific management strategies. We are exploring it's clinical utility further by adding more data from a tertiary cancer hospital. We have also deployed a patented prototype software for research use to study the effect of integrating this model in clinical workflows to optimize patient care.