This paper introduces CEET-Fed - a framework that combines three cutting-edge concepts in AI: federated learning (training across distributed sites without sharing patient data), an ensemble of CNN models (VGG16 and ResNet50) plus a Vision Transformer (ViT), and LIME-based explainability for interpreting predictions.
The system predicts breast cancer BI-RADS categories (a standardized imaging classification system) and molecular subtypes from a set of 41 clinical risk factors rather than from imaging directly. The risk factors include family history, menstrual history, reproductive history, lifestyle factors, and prior biopsies.
Published in Scientific Reports in 2025, this study addresses a significant need: many women at high risk for breast cancer lack access to high-quality imaging centers, but clinical risk data is widely collectable through routine consultation.
Training AI models requires large amounts of data, but medical data is subject to strict privacy regulations. Patient records cannot simply be pooled across hospitals without extensive consent and legal frameworks. Federated learning solves this problem: instead of sharing patient data, only the model weights (the learned patterns, not the data) are shared between institutions.
Each participating hospital trains the model on its own local data, then sends only the updated model weights to a central server that aggregates them. The patient data never leaves each hospital. This makes federated learning a privacy-preserving alternative that still enables large-scale collaborative model training.
The study used data from the National Cancer Control Foundation in Yemen - a setting where access to AI and advanced diagnostics is limited. This context is important: the model is designed for resource-constrained environments where imaging-based screening is not always feasible.
CEET-Fed uses an ensemble approach at its core: VGG16 and ResNet50, two powerful CNN architectures originally designed for image recognition, are adapted to process tabular risk factor data. Their outputs are combined with predictions from a Vision Transformer (ViT) - an architecture that uses attention mechanisms to identify relationships between different risk factors.
Vision Transformers were originally developed for image analysis but have shown strong performance on structured (tabular) data as well. By treating each clinical risk factor as a 'token' (similar to how words are processed in language models), ViT can capture complex interactions between risk factors that simpler models miss.
The three model outputs are fused through a final classification layer. This ensemble approach reduces the risk that any single model's errors dominate - if one model is uncertain or wrong, the others can compensate. The federated training ensures each model variant learns from all participating institutions.
The study used 734 patients from the National Cancer Control Foundation in Yemen, with 41 clinical risk factors collected per patient. The dataset is relatively small, which is acknowledged as a limitation, but the federated learning approach is designed to be extended to larger multi-site datasets.
The model was trained to predict three different classification targets: binary (benign vs. malignant), multi-class molecular subtypes (Luminal A, Luminal B, HER2-enriched, Triple-negative), and BI-RADS categories (a 6-level classification from Category 0 to Category 6 used in breast imaging reporting).
BI-RADS prediction from risk factors alone - without imaging - is a novel contribution. Being able to estimate a patient's BI-RADS category from clinical data could help prioritize imaging resources for those most likely to have significant findings.
In the federated learning setting, CEET-Fed achieved 98.65% accuracy for binary classification, 97.30% for multi-class molecular subtype classification, and 95.59% for BI-RADS classification. These are remarkably high figures for complex multi-class tasks using only clinical risk factors.
The federated model performance was close to (within 1-2%) the performance of a centrally trained model that had access to all data simultaneously. This is the key validation of federated learning - it should not significantly degrade performance compared to centralized training.
The model outperformed standalone VGG16, standalone ResNet50, and standalone ViT on all three tasks, confirming that the ensemble combination provides meaningful performance gains over any individual component.
LIME (Local Interpretable Model-agnostic Explanations) was applied to interpret individual predictions. LIME works by perturbing the input (slightly changing individual risk factors) and observing how predictions change, building a locally interpretable explanation for each case.
Key risk factors identified by LIME as strongly influential include family history of breast cancer, age at first menstruation, age at first pregnancy, and prior biopsy history. These align well with established epidemiological risk factors for breast cancer.
Crucially, LIME explanations can be presented to patients and clinicians in plain language: 'Your prediction was primarily driven by your family history and your early onset of menstruation.' This transparency increases trust and enables clinical sense-checking of AI recommendations.
CEET-Fed's reliance on clinical risk factors rather than imaging is a deliberate design choice for contexts where imaging infrastructure is limited or costly. A questionnaire-based tool that estimates BI-RADS category and molecular subtype probability could guide which patients most urgently need imaging referral.
In Yemen and similar settings, where mammography screening is not yet universal, this kind of AI-assisted triage could direct limited imaging resources to the highest-risk patients, potentially improving early detection rates without requiring immediate scaling of imaging infrastructure.
The federated learning framework also means the model could be deployed across multiple community health centers and district hospitals, each training locally on their patient populations and contributing to a shared global model, without any center compromising patient privacy.
CEET-Fed demonstrates that state-of-the-art AI architectures - CNNs, Vision Transformers, and explainability tools - can be deployed in a privacy-preserving federated framework without sacrificing prediction accuracy.
The combination of clinical risk factors (widely available), federated training (privacy-preserving), and LIME explanations (trustworthy) makes this framework particularly suited for deployment in healthcare systems with limited infrastructure and strong privacy concerns.
Future work should expand to larger multi-national datasets, incorporate longitudinal risk factor data, and explore integration with imaging data when available - potentially enabling a hybrid risk-plus-imaging model for the highest-risk patients.