Elderly health requires continuous monitoring to enable early detection of neurodegenerative conditions. This study positions Parkinson’s disease (PD) detection as a core function of an IoT (Internet of Thing)-based elder monitoring system that collects voice and sensor data for remote analysis. Initially, voice signals are denoised using an adaptive wavelet thresholding (AWT) method, which effectively suppresses background noise and enhances the image. The proposed PD-LSTM can be integrated as the deep learning decision module in an IoT-based elder monitoring framework, enabling automated, continuous monitoring and alerting for caregivers and clinicians. Mel Frequency Cepstral Coefficients (MFCC) are used as a feature extraction technique to produce discriminant features, and a sparse autoencoder is used to extract the features of the voice signal (VS). Finally, the Bi-directional LSTM (BDLSTM) used to classify the PD such as normal, and Parkinson. The proposed PD-LSTM approach not only enhances Parkinson’s detection accuracy but also forms a potential component of an IoT-enabled elder monitoring ecosystem, providing continuous and intelligent healthcare assistance. The performance of the PD-LSTM approaches was assessed using the metrics such as F1 score, specificity, recall, accuracy, and precision. The PD-LSTM approach achieves a high accuracy of 99.22% for PD. The PD-LSTM improves the accuracy range of 10.47%, 3.19% and 11.85% better than ZFNet-LHO-DRN, FB-DNN, and Ma-ST-DGN respectively