Development and Evaluation of Enhanced Machine Learning Models for Tumor Severity Classification Through Integrated Feature Extraction and Selection Techniques

Abstract

The precise categorisation of the severity of tumours is an important function in medicine, as it is a vital part of deciding treatment and predicting patient outcome. We offer an integrated pipeline for developing and evaluating the latest machine learning methods for tumour severity classification. We apply advanced methods for feature extraction and selection to improve the model’s accuracy, interpretability, and robustness. Ultimately, the framework organises diverse biomedical datasets for later analysis through domain-specific preprocessing approaches. Methods such as PCA, wavelet transform, and texture features are used to extract important information from raw data, thereby reducing dimensionality and preserving diagnostic information. More advanced techniques for feature selection are then used, such as Recursive Feature Elimination (RFE). Moreover, mutual information, which identify the most relevant features in the input dataset to reduce computational cost and the risk of overfitting. The models are validated using cross-validation techniques, with accuracy, precision, recall, F1-score, and Area Under the Receiver Operating Characteristic (ROC) Curve (AUC-ROC) as metrics. The results demonstrate that combining feature extraction and selection boosts classification accuracy and model generalisation across multiple datasets. Furthermore, explainability analysis clarifies the contribution of selected features to model predictions, which is crucial for building trust and transparency in clinical applications. However, really, what it demonstrates is the promise of all these things in concert. State-of-the-art machine learning, combined with rigorously designed features, will ultimately address some of those more difficult diagnostic challenges. These findings also pave the way for scalable, interpretable and clinically applicable AI-based solutions which may help in cancer detection and management. In this work, we will extend the proposed pipeline to other medical imaging modalities and explore real-time implementations.