Breast cancer multidisciplinary team (MDT) meetings are central to evidence-based oncology, yet manual deliberation processes suffer from inconsistency, time pressure, and limited reach in low-resource healthcare settings. This paper presents an AI-driven breast cancer treatment planning system trained on 1,000 real-world MDT records collected from health facilities in Ghana. Two multi-output machine learning classification models were developed to predict up to 15 structured MDT treatment decisions simultaneously. Model1 employs nine focused radiological and pathological input features, achieving a mean test macro-F1 score of 0.5488 and an average classification accuracy of 63.6% across 12 output targets, using a Multi-Output Classifier with a balanced Random Forest backend. The Model 2 extends the feature space to 94 structured variables drawn from radiology, pathology, and surgical reports and predicts 15 MDT outputs using Gradient Boosting, achieving a mean macro-F1 of 0.6632 and an average accuracy of 75.8%, a substantial improvement that can be attributed to the inclusion of receptor status and molecular subtype data. Both models are deployed through an interactive Gradio web application hosted on Hugging Face to provide clinicians with structured predictions and natural-language clinical explanations. The system demonstrates that structured MDT outputs including treatment intent, chemotherapy regimen, surgical plan, radiotherapy targets, and endocrine therapy type can be reliably predicted from routine clinical data. This work represents a reproducible and accessible AI decision-support framework with direct relevance to breast cancer care in sub-Saharan Africa and other resource-constrained healthcare settings.
CORE (COnnecting REpositories)
CORE (COnnecting REpositories)