Early detection of Parkinson’s disease (PD) is crucial for prompt intervention and the best possible therapeutic management. In Parkinson’s disease datasets, the number of healthy samples exceeds the number of disease-positive cases. This scenario is called class imbalanace. In this paper, we explore the application of data balancing techniques. In order to improve classification accuracy and refine the feature space, we also employ sophisticated feature-selection techniques. By integrating these techniques, we seek to construct a more balanced and robust classification framework employing CNN-LSTM and Feedforward Neural Network (FNN) architectures to accurately identify Parkinson’s disease cases. Experimental results show that the proposed approach alleviates class imbalance, facilitates feature selection, and augments the model sensitivity to early stage Parkinson’s disease, demonstrating its potential as a useful tool in clinical diagnosis and decision making.