When breast cancer spreads, it typically moves first to the sentinel lymph node - the first lymph node in the drainage pathway from the tumor. Determining whether cancer has reached the sentinel node is critical for staging the disease and planning treatment.
Currently, this is determined through sentinel lymph node biopsy (SLNB), a surgical procedure that removes and examines the node under a microscope. While generally safe, SLNB carries risks including arm swelling (lymphedema), infection, and numbness.
If an AI model could reliably predict whether the sentinel node is cancerous before surgery - using only information already available from clinical assessments - it could help identify patients who can safely skip the biopsy procedure.
The study included 1,832 breast cancer patients treated at Shiraz University of Medical Sciences in Iran. For each patient, clinical and pathological features collected before surgery were used as model inputs.
The AI model used was TabNet - a deep learning architecture specifically designed for tabular data (structured tables of features), as opposed to images or text. TabNet uses sequential attention to select which features are most relevant for each individual prediction.
Multiple machine learning approaches were compared: TabNet, Random Forest, logistic regression, support vector machine (SVM), and others. This allowed the researchers to determine whether the deep learning approach offered meaningful advantages over simpler methods.
TabNet processes tabular data through sequential attention steps. At each step, the model learns to focus on the subset of input features most relevant for making the prediction, and combines information across steps to arrive at a final output.
This attention mechanism has a key advantage for clinical AI: it provides a natural way to identify which features the model relies on most, making predictions more interpretable than a black-box approach.
The input features included: vascular invasion (whether cancer cells were found in blood or lymph vessels), tumor size, core needle biopsy pathology results, patient age, and family history of breast cancer, among others.
TabNet achieved an AUC (area under the ROC curve) of 0.74, meaning it correctly ranked positive cases (lymph node metastasis present) above negative cases 74% of the time. This was among the best performances of all models tested.
The most important predictor identified by the model was vascular invasion - the presence of cancer cells in blood or lymph vessels near the primary tumor. This is a well-established clinical risk factor, and the model's ability to identify it as most important validates its clinical reasoning.
Other top features included tumor size, pathology results from the core needle biopsy, patient age, and family history of breast cancer. This list aligns well with clinical knowledge about sentinel node metastasis risk factors.
For patients predicted to have a very low probability of sentinel node metastasis, this model could support a clinical decision to forego SLNB. However, the AUC of 0.74 means the model is not perfect, and clinical judgment would still be needed to weigh the risks.
The study was conducted entirely in an Iranian patient population, which may differ from Western populations in tumor biology, clinical presentation, and genetic background. External validation in other populations is needed before the model can be broadly applied.
Access to SLNB can be limited in lower-resource healthcare settings. In these contexts, an AI prediction tool could be particularly valuable for triaging patients toward or away from the procedure.
This study demonstrates that TabNet - a deep learning model designed for tabular clinical data - can predict sentinel lymph node metastasis in breast cancer patients using preoperative features with reasonable accuracy.
The interpretability of the model's feature selection is an important clinical benefit. Knowing that the model relies on vascular invasion and tumor size - known clinical risk factors - builds confidence in its reasoning and facilitates clinical adoption.
Future work should focus on external validation in diverse populations, integration of imaging biomarkers, and prospective evaluation in clinical settings where the model's predictions would be used to inform actual biopsy decisions.