Visual, Auditory, and Kinesthetic (VAK) learning styles fundamentally shape how students engage with and comprehend educational content, yet their subject-specific impacts remain understudied in diverse educational contexts. Despite widespread recognition of learning style importance, limited research examines how VAK preferences influence student achievement across different subjects, particularly in South Asian secondary education systems where traditional one-size-fits-all teaching approaches dominate. This study investigated VAK learning style impacts on student performance across six core subjects (Mathematics, Science, English, History, Religion, and Mother Language) in Sri Lankan secondary schools using advanced ensemble machine learning techniques. We collected data from multiple sources including Kaggle’s Student Learning Preferences dataset and custom surveys, then applied three ensemble methods bagging, boosting, and stacking to classify learning styles and analyze their subject-wise effects. The stacked ensemble model, utilizing Feedforward Neural Networks and Random Forest as base learners with XGBoost as meta-learner, achieved 91.70% classification accuracy. Results revealed distinct learning style preferences by subject: kinesthetic dominance in Mathematics (41.4%) and Science (47.8%), auditory preference in Mother Language (40.9%) and History (37.5%), and balanced distributions in English and Religion. These findings demonstrate that optimal learning styles vary significantly across subjects, challenging uniform pedagogical approaches. The research provides actionable insights for educators to adapt teaching methods according to subject-specific learning style distributions, potentially improving student engagement and achievement. This work establishes a foundation for developing adaptive learning systems that can dynamically adjust instructional strategies based on both individual learning preferences and subject requirements.