Heart disease is a major health risk for people worldwide. It is a complex medical condition that affects the cardiovascular system. Predictive techniques have been developed and implemented to detect heart disease. Several factors, including unhealthy lifestyle and lack of physical activity, play a critical role in diagnosing the disease. There are two approaches for predicting heart disease: physical and chemical modeling and artificial intelligence (AI). Mathematical models rely on complex physics and chemistry concepts to create intricate models that increase in complexity as the level of detail increases. AI techniques, on the other hand, can automatically generate simpler models based on raw measurements using both deterministic and heuristic methods. However, physical modeling must strike a balance between complexity and assumptions to create simplified yet representative models. Feature selection involves identifying the most significant variables in a model, known as features. The procedure is supported by statistical and search tools that facilitate experts in examining extensive datasets and identifying the most significant characteristics. This research utilizes a public dataset derived from four databases, which originally contained 76 attributes. Including the predicted attribute. However, we focus on 14 attributes including the target as most published experiments have done. The method we propose is to use pattern recognition to improve the accuracy of a model by identifying important correlations and eliminating features with insufficient variance. This simplifies the training process and reduces complexity, resulting in better results than models without feature selection. To strengthen our findings, we employed several filtering methods based on statistical metrics for feature selection, such as mutual information gain, Chi-square, Fisher score, Pearson correlation, and variance thresholding. In addition, we evaluate wrapper methods based on greedy search, namely forward search, backward search and exhaustive search, with the possibility of selecting up to six features. We evaluate the accuracy of the models by using Naive Bayes classifiers, Neyman Pearson Criterion and Hidden Markov Models and thus perform a detailed comparison of the models’ accuracies. Finally, after performing the analysis using the filter, wrapper and classifier methods we have found that the most effective way to predict heart disease is with a Naïve Bayesian classifier, using three specific features, age, chest pain and gender. This particular configuration has resulted in an accuracy rate of 87%.