Forecasting-based biomedical time-series data synthesis for open data and robust AI.

Area of Science:

• Biomedical data science
• Artificial intelligence in healthcare
• Time-series analysis

Background:

• Biomedical time-series AI development is hindered by data scarcity due to privacy regulations and resource intensity.
• A gap exists between the data needs of AI models and the accessibility of real-world biomedical data.
• Existing synthetic data methods like GANs and VAEs capture global distributions but may not be optimal for sequential data.

Purpose of the Study:

• To propose a novel framework for generating synthetic biomedical time-series data.
• To leverage advanced forecasting models for high-fidelity replication of complex electrophysiological signals (EEG, EMG).
• To enable open AI development by providing shareable, privacy-preserving synthetic datasets.

Main Methods:

• Developed a synthetic data generation framework utilizing recent forecasting models.
• Focused on capturing the sequential dynamics inherent in biomedical time-series data.
• Validated the framework on electroencephalography (EEG) and electromyography (EMG) signal generation.

Main Results:

• The proposed framework accurately replicates complex electrophysiological signals with high fidelity.
• Synthetic data augmentation improved downstream sleep-stage classification performance by up to 3.71%.
• A model trained solely on synthetic data achieved 91.00% accuracy, outperforming a real-data-only baseline.

Conclusions:

• Forecasting models provide an effective approach for generating high-fidelity synthetic biomedical time-series data.
• Synthetic data generated by this framework can significantly enhance AI model performance and facilitate open research.
• This method addresses data accessibility challenges while preserving patient confidentiality.