Arabic medical text classification is a particularly challenging task rooted in the linguistic complexity of the Arabic language and the scarcity of annotated medical datasets. This study investigates advanced transfer learning techniques by fine-tuning pre-trained models, namely AraBERT and AraELECTRA and introducing a novel hybrid architecture that integrates AraBERT with a Transformer Encoder and a Multi-Layer Perceptron. Interestingly, the results reveal that while architectural enhancements offer theoretical promise, they do not consistently lead to performance improvements. Instead, the most significant improvements were achieved by applying robust data preprocessing and strategic class restructuring. In addition, efficient data cleaning, increasing data size, and well-curation of the training sets appear to be key factors for performance optimization. These findings prove that the quality and organization of training data often outweigh the potential of innovative architectures in achieving high-performance levels in Arabic medical text classification. This study emphasizes the vital role of preprocessing and class distribution strategies in advancing Arabic NLP applications and data challenges in healthcare.