Feature Selection Using A Hybrid Approach: Harmony Search with Recurrent Neural Networks and Filter Methods on the Mimic-III Dataset

Building reliable and effective prediction models requires careful feature selection, especially when working with complicated medical datasets. Applied to the Medical Information Mart for Intensive Care III (MIMIC-III) dataset, this study presents a novel hybrid method for feature selection that incorporates Harmony Search (HS), Recurrent Neural Networks (RNNs), and conventional Filter Methods. A wide range of clinical records, including vital signs, prescriptions, test results, and diagnostic codes, are included in the dataset. With the help of filter techniques, we guide the first feature ranking in our proposed method (HS-RNN-FM), which combines the deep learning characteristics of RNNs with the global optimization capabilities of Harmony Search. According to experimental data, HS-RNN-FM performs better in terms of precision, recall, accuracy, F1-score, and AUC-ROC than conventional techniques like Genetic Algorithm with Mutual Information (GA-MI) and Particle Swarm Optimization with Chi-Square Test (PSO-Chi2). Due to its exceptional performance, HS-RNN-FM has the potential to improve predictive modeling in healthcare applications by handling huge and complicated datasets in an efficient manner.