TSG-DDT: Time-Series Generative Denoising Diffusion Transformers

Deep Brain Stimulation (DBS) involves implanting electrodes in the brain to modulate neural activity, treating conditions like Epilepsy [1]. Using synthetic micro-electrode array (MEA) signals as stimuli in a DBS system enables electrical impulses that mimic natural brain activity, improving precision and effectiveness [2]. This can reduce side effects and enhance therapeutic outcomes [3]. This study introduces Time-Series Generative Denoising Diffusion Transformers (TSG-DDT), an architecture for generating high-fidelity synthetic MEA data [4]. TSG-DDT combines denoising diffusion models [5] with transformer blocks [6] to overcome limitations in existing MEA data generation techniques. TSG-DDT uses a time-dependent sequential noise schedule to systematically add noise with a cosine schedule [7] to clean MEA data. This method ensures that all points in the sequence follow the same schedule independently. The model is trained to extract noise from the noisy signal, making the original signal easily reconstructable. The architecture includes a tokenizer to embed patches of the sequence into a latent space using a feed-forward network (FFN) [8], transformer encoder layers [9], and an FFN-decoder to map back to the original space. The model’s performance was evaluated using Marginal Distribution Difference (MDD), AutoCorrelation Difference (ACD), Skewness Difference (SD), Kurtosis Difference (KD), Euclidean Distance (ED), and Dynamic Time Warping (DTW) [10]. Synthetic MEA data generated by TSG-DDT was compared to both training and test sets, and the training and test sets were compared for evaluation. Results showed MDD values of 0.084 (Training vs Synthetic), 0.086 (Test vs Synthetic), and 0.032 (Training vs Test). ACD values were 0.1613, 0.1771, and 0.1555, respectively. SD values were 7.2793, 8.417, and 4.059. KD values were 179.557, 283.675, and 174.217. ED values were 16.572, 15.430, and 14.718. DTW values were 107.582, 101.630, and 96.966, respectively. These metrics indicate the fidelity and realism of the generated data, demonstrating that the synthetic data closely resembles real MEA data, with marginal differences across all metrics, underscoring the model's effectiveness. This opens the possibility of using TSG-DDT for biomimetic brain stimulation.