GAN and LSTM-based collaborative tremor classification approach for next generation healthcare system

To adopt different deep learning (DL) techniques for Essential tremor (ET) and Parkinson's tremor (PST) classification, a collaborative approach to address the misdiagnosis of healthy and PSD patients based on ET and PST's frequency patterns and severity.

This classification uses a comprehensive PDBioStamp time-series dataset to classify tremors based on action and rest tremors of healthy and post-stroke depression (PSD) patients. Combination of DL models, including generative adversarial network (GAN) to generate synthetic data, Autoencoder to reduce the size dimensionality of data and learn a latent representation, long short-term memory (LSTM) to capture temporal features and essential characteristics to improve the performance of tremor classification.

The performance of the models is evaluated using evaluation metrics, such as training and testing accuracy, F1 score, loss by varying different epochs, and different optimizers. Moreover, the performance of the model was compared using different state-of-the-art works.

The ET and PST classification result indicates that the proposed combination of GAN, autoencoder, and LSTM outperformed with 80.0 training, 80.3 testing accuracy, 0.82 F1 score, and 0.89 AUC, which is higher than existing DL models. The proposed collaborative approach helps doctors improve their diagnosis for ET and PST tremor patients. This classification helps doctors to identify PSD and healthy control patients.