This paper explores the enhancement of predictive models using Cellular Automata (CA) to create neighborhood-influenced features in machine learning models. Diabetes prediction models are crucial in early diagnosis, allowing timely intervention and management. We investigated the performance of four techniques: LightGBM, XGBoost, Random Forest, and Logistic Regression, both with and without CA-based feature engineering on a diabetes dataset. After encoding feature values into discrete states, CA principles were applied to generate new features that capture risk levels based on the influence of \( k \) closest neighbors. A comparative study of model accuracy, precision, recall, and the F1 score shows that CA-enhanced models provide better insights into neighborhood-driven predictions. This paper illustrates the potential for CA to improve traditional machine-learning techniques for classification tasks in diabetes prediction.