level: research
longitudinal tumor measurements, dropout times, and genetic data each offer clues about treatment response, but combining them in one model is hard. this work extends the empirical bayes variational autoencoder (eb-vae) to handle both tumor growth over time and time-to-dropout events. the model uses latent individual effects to capture differences between patients, with a prior that depends on genetic covariates. a decoder then turns these latent effects into tumor volume predictions.
to handle informative dropout, where patients leave the study for reasons related to their disease, the decoder includes a hazard model. this lets the model predict both tumor trajectories and dropout times together. the authors tested two decoder types: a fully neural network and a hybrid semi-mechanistic one that blends known biology with learned parts. they also added genomic information by conditioning the prior on genetic covariates.
the hybrid decoder showed promise in early results, but full performance details are not yet available. the approach aims to give a more complete picture of patient outcomes by linking tumor dynamics and dropout risk. it could help in clinical trials where missing data from dropouts can bias results. the method builds on existing eb-vae ideas and adapts them for survival data.
why it matters: better joint models can improve treatment effect estimates in clinical trials by accounting for patients who drop out due to disease progression.