Article Analysis

Our approach involves a multi-stage process: first, fine-tuning an Arabic BERT model with medical data for vector representations; second, applying PCA and ICA for dimensionality reduction and source separation; third, incorporating modified loss functions with information entropy for enhanced certainty; and finally, training an enhanced Arabic BERT model on specialized medical text. Data augmentation techniques like back-translation were used to balance the imbalanced dataset, and k-fold cross-validation ensured robust evaluation.

The enhanced PCA-based model achieved 94.7% accuracy, surpassing human-level performance of 94.3%. Data augmentation significantly improved model stability, with back-translation yielding the best results. PCA and ICA effectively enhanced data separation and reduced dimensionality. The entropy model showed smoother probability transitions, and the Medical Enhanced BERT model demonstrated superior overall performance and confidence calibration, especially for correct predictions. These findings offer valuable tools for policymakers and researchers to ensure trustworthiness in Arabic online health information.

The study acknowledges limitations, including dataset imbalance (skewed towards low-quality pages) and the fixed 512-token document size for BERT, which might lead to information loss. Future work will focus on addressing dataset imbalance with more diverse data, exploring alternative models like Longformer for longer input sequences, developing more robust AI safety mechanisms, and validating models with class-wise temperature scaling and external datasets for broader generalizability and real-world integration into search engines and health chatbots.