Abstract

Blood cancer, encompassing leukemia, lymphoma, and multiple myeloma, remains among the most life-threatening hematological malignancies globally. Early and accurate diagnosis is paramount for improving survival rates. This paper introduces ALD-RIO (Adaptive Learning with Deep-Residual Iterative Optimization), a novel hybrid deep learning architecture designed for multi-class blood cancer disease prediction. The proposed framework integrates Bidirectional Long Short-Term Memory (BiLSTM) networks with Residual Network (ResNet) blocks and a multi-head self-attention mechanism, guided by Adaptive Lasso Feature Selection (ALFS) to identify the most discriminative biomarkers from gene expression, Complete Blood Count (CBC) parameters, and clinical features. The model employs SMOTE-based class balancing, Bayesian hyperparameter optimization, and an iterative residual refinement strategy to address class imbalance and overfitting challenges. Extensive experiments on five publicly available benchmark datasets demonstrate that ALD-RIO achieves 97.83% accuracy, 97.45% precision, 97.91% recall, 97.68% F1-score, and 0.989 AUC-ROC, outperforming existing state-of-the-art methods by a significant margin of 3.62% in accuracy. Ablation studies confirm the contribution of each architectural component. The proposed model holds strong potential for clinical deployment in hematology diagnostics.

Authors

T. Nikil Prakash
St. Joseph’s College, India

Keywords

Blood Cancer Prediction, ALD-RIO, Deep Learning, BiLSTM, Attention Mechanism, Gene Expression, Hematological Malignancy, Feature Selection