Denoising diffusion model for increased performance of detecting structural heart disease.

Recent advancements in generative artificial intelligence have shown promise in producing realistic images from complex data distributions. We developed a denoising diffusion probabilistic model trained on the CheXchoNet dataset, encoding the joint distribution of demographic data and echocardiogram measurements. We generated a synthetic dataset skewed towards younger patients with a higher prevalence of structural left ventricle disease. A diagnostic deep learning model trained on the synthetic dataset performed comparably to one trained on real data producing an AUROC=0.75(95%CI 0.72-0.77), with similar performance on an internal dataset. Combining real data with positive samples from the synthetic data improved diagnostic accuracy producing an AUROC=0.80(95%CI 0.78-0.82). Subgroup analysis showed the largest performance improvement across younger patients. These results suggest diffusion models can increase diagnostic accuracy and fine-tune models for specific populations.