Stabilizing machine learning for reproducible and explainable results: A novel validation approach to subject-specific insights