Parkinson Disease (PD) is a progressive neurodegenerative disease characterized by different tremor patterns making it difficult to diagnose and to follow. The proposed research will include a hybrid deep machine learning approach that will integrate Autoencoder-based nonlinear feature extraction with Gradient Boosting classification to detect accurate tremor patterns. Triaxial accelerator signals were sorted into raw triaxial signals, and Butterworth Band-Pass Filter (2.512.5 Hz) was used to preprocess raw triaxial signals to isolate tremor frequencies. Based on the filtered signals, 92 time- and frequency-domain features were obtained using the statistical-based, spectral-based and principal component-based techniques. A Gradient Boosting classifier was then trained on the latent nonlinear features together with extracted descriptors, which were learned on an Autoencoder and used to distinguish between rest, postural, kinetic, and constancy tremors. The hybrid model had better accuracy and strength in comparison to traditional approaches. This combined system evidences the possibilities of using the combination of deep and ensemble learning in the objective evaluation of tremors and the initial diagnosis of Parkinson Disease.