Researchers optimize variational autoencoders to better predict survival risk in multiple myeloma by balancing latent space regularization with prognostic accuracy.

MO-RiskVAE extends the MyeVAE framework to integrate multiple data types (omics and clinical) for improved survival prediction in multiple myeloma patients

Study reveals that standard latent regularization in VAEs fails to preserve disease-relevant information when trained with survival supervision, causing unstable predictions

Systematic investigation isolates three critical latent design factors: regularization scale, posterior geometry, and latent space structure to understand which elements drive model performance

Findings demonstrate that survival-driven VAE training is particularly sensitive to the magnitude and structural properties of the latent space representation