Attention-Based Deep Learning Framework for Multi-Class Classification of Liver Cirrhosis Stages Using Clinical and Biochemical Biomarkers

Abstract

ABSTRACT liver cirrhosis staging is crucial to enhance prognosis among patients and optimize the therapeutic approach. An attention-based deep learning framework for automated cirrhosis stage classification using routinely collected clinical and biochemical features As solution 1: A dataset of 25,000 patient records with 19 medical attributeswas prepared after preprocessing missing-value imputation with random-forest, label encoding, and feature standardization and development model. By applying dense neural layers combined with a custom attention mechanism, we get an architecture capable of highlighting the most discriminative features while diminishing less informative patterns. On an independent test set, predictive performance (accuracy = 90.46%, Weighted F1-Score = 0.9048), and multi-class ROC analysis (AUC range = 0.970–0.984 for all stages) demonstrated that we possess high predictive power. A 5-fold cross-validation procedure further validated the method with a mean accuracy of 0.9124 ± 0.0032 demonstrating robustness and generalizability. Conclusion These results suggest that the attention model proposed is a strong, interpretable, and clinically applicable model for classifying liver cirrhosis stages and may be applied to computer-aided diagnosis systems and clinician-targeted risk stratification of patients. Keywords: Liver Cirrhosis Classification, Attention-Based Deep Learning, Clinical Biomarkers Analysis, Computer-Aided Diagnosis