Research and development of highly accurate falling detection systems (FDSs) for individuals with medical conditions or the elderly are crucial for mitigating the risks associated with falls. These systems are of great significance for real-world applications in healthcare and elderly care. However, improving the accuracy of the FDS is a significant challenge, particularly due to the difficulty in collecting comprehensive and accurate fall data which causes imbalance in datasets. In this paper, we introduce a Genetic Algorithm (GA) based Random Forest (RF) method, named GA4RF, to enhance the accuracy of fall detection models by optimizing their hyperparameters. Specifically, we propose a fitness function based on the Matthews Correlation Coefficient (MCC) to enable the GA to identify the optimal hyperparameter set for the RF classifier. This targeted optimization ensures that the algorithm prioritizes balanced accuracy across both fall and non-fall events, mitigating the bias inherent in imbalanced datasets. Additionally, we select a chromosome structure focusing on three hyperparameters to effectively narrow the search space, reducing computational complexity while maintaining high detection performance. Experimental evaluations on the MobiAct v2.0 and UP-Fall detection datasets demonstrate that GA4RF achieves high human activity recognition accuracy, especially a significant improvement of fall detection performance. Compared to the conventional RF method, the F1-score increases from 95.23% to 96.91% on the MobiAct v2.0 dataset and from 96.16% to 97.67% on the UP-Fall dataset. These results indicate that GA4RF is a promising approach for improving FDS, especially when dealing with highly imbalanced data and meeting performance requirements in real-world applications.