<p>Approximately 30% of all deaths worldwide result from heart disease, according to a recent study. It is common for doctors to make diagnoses and treat patients based on their experience and knowledge, but occasionally incorrect diagnoses and treatments are documented. A doctor may recommend several costly and time-consuming diagnostic procedures to identify the exact origin of the patient’s illness and provide the best possible care. Due to the size of medical databases, quick processing is generally not possible. This situation calls for the use of machine learning methods. Based on the recorded symptoms of a patient, the paper proposes a stacking framework that uses harmony search optimization (HSO) to predict early cardiac disease in the course of the disease diagnosis process. There are two stages in the framework: a stacking stage, and a base-learners’ combination stage that utilizes optimization. By combining the base learners at the base layer with HSO, the performance of the model is optimized. Open national government machine learning repository (NGML) and Cleveland datasets are used as test datasets for evaluating model effectiveness. In contrast to base models and other current research, the overall accuracy attained in single layer stacking is 87%, multilayer stacking is 89%, and the stacking model employing HSO is 92%.</p>