Synthetic Data Generation for AI in Medicine

The difficulty of accessing real-world data in the clinical sector represents a significant obstacle to the development of effective artificial intelligence models for diagnostics and prevention. Generative AI offers a promising solution to increase the amount of data and improve model training, as already demonstrated in computer vision and natural language processing (NLP).

TransConv-DDPM: A New Approach

A new study introduces TransConv-DDPM, an advanced generative AI method for creating biomechanical and physiological data in the form of time series. The model is based on a denoising diffusion probabilistic model (DDPM) with U-Net, multi-scale convolution modules, and a transformer layer. This architecture allows capturing both global and local temporal dependencies.

Promising Results

TransConv-DDPM was evaluated on three different datasets, generating both long and short time series. The results, compared with state-of-the-art methods such as TimeGAN and Diffusion-TS, show promising performance, particularly on the SmartFallMM and EEG datasets. The addition of synthetic data generated by TransConv-DDPM improved the performance of a predictive model on the SmartFallMM dataset, with a 13.64% increase in F1-score and a 14.93% increase in overall accuracy compared to the baseline model trained only on real data.