Interpretable hybrid three tier LSTM model for accurate and transparent breast tumor classification in clinical decision support

The accurate classification of cancerous tumors is essential for early diagnosis and optimal treatment planning. However, traditional machine learning methods often struggle to model the complex and high-dimensional nature of biomedical data. Even though deep learning models have recorded significant improvements, they frequently lack interpretability, limiting their clinical trust and adoption. We propose a Hybrid 3-Tier of Long Short-Term Memory (LSTM) Model (where “3-Tier” denotes functional stages, not the number layers) for breast tumor classification. The proposed model integrates sequential learning with classical machine learning. It is designed to simultaneously address challenges of high dimensionality and limited transparency in existing literature. We employed the Synthetic Minority Oversampling Technique (SMOTE) to handle class imbalance. For our model explainability and interpretability, we used SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) to interpret predictions at the feature level. Our model performance was evaluated on a publicly available breast cancer dataset from Kaggle. The proposed Hybrid 3-Tier LSTM model obtained a F1-score of 97%, AUC of 0.997 with high precision and recall for both benign and malignant classes. Our LIME and SHAP analysis revealed clinically relevant features that aligned with known tumor characteristics. We achieved model interpretability without compromising predictive performance. Our results show the feasibility and efficacy of hybrid deep learning approaches for interpretable cancer classification. Our model offers a clinically meaningful balance between accuracy and transparency. Thus, the proposed model is well-suitable for integration into decision support systems in oncology.