Cardiovascular conditions are among the primary reasons of death in the developed world. Cardiovascular conditions often go undetected for long periods until severe conditions develop. Early diagnosis and timely treatment have been shown to significantly improve outcomes for patients. Heart electrical signals can be indicative of underlying and developing conditions and can be monitored by an electroencephalogram (EEG). However, interpreting EEG signals is a challenging task that requires significant training and experience. The paper investigates the use of recurrent neural networks for the detection of heart conditions. Signals squired from EEG are treated as a multivariate time series and networks are tasked with detecting anomalous signals. To improve network performance, hyperparameter optimization is performed using a modified metaheuristic introduced for this work. Several optimizers were compared to the introduced algorithm under identical test conditions on an authentic publicly available dataset. The modified algorithm outperforms competitors including the original version of the algorithm evaluated in both objective function (error rate) and detailed metrics such as accuracy, precision, recall and f1-score. The best performing optimized model attained accuracy of 99.26% a better outcomes compared to models optimized by other evaluated contemporary metaheuristics.