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

Area of Science:

• Neurology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Essential tremor (ET) and Parkinson's tremor (PST) share similar frequency patterns, leading to misdiagnosis.
• Accurate classification is crucial for effective treatment and differentiating from other conditions like post-stroke depression (PSD).

Purpose of the Study:

• To develop and evaluate a collaborative deep learning (DL) framework for classifying ET and PST.
• To address misdiagnosis by analyzing tremor frequency patterns and severity.
• To distinguish between healthy individuals, ET, PST, and PSD patients.

Main Methods:

• Utilized a PDBioStamp time-series dataset for classifying action and rest tremors.
• Employed a combination of DL models: Generative Adversarial Network (GAN) for synthetic data, Autoencoder for dimensionality reduction, and Long Short-Term Memory (LSTM) for temporal feature extraction.
• Evaluated model performance using accuracy, F1 score, and AUC, comparing against state-of-the-art methods.

Main Results:

• The combined GAN-Autoencoder-LSTM model achieved 80.0% training accuracy and 80.3% testing accuracy.
• The model obtained an F1 score of 0.82 and an AUC of 0.89.
• Performance metrics surpassed existing deep learning models for tremor classification.

Conclusions:

• The proposed DL approach demonstrates superior performance in classifying ET and PST.
• This classification system aids clinicians in improving diagnostic accuracy for tremor patients.
• The framework effectively assists in identifying patients with PSD and distinguishing them from healthy controls.