#12 · ML / AI methods

Generative models for patient care trajectories in continuous time

The shared thread is generative modeling of clinical event sequences in which timing is represented explicitly rather than discretized. The Transformer Hawkes process models patient care trajectories in continuous time and addresses rare events with inverse square-root class weighting. Multiplicity-weighted Stochastic Attention, a Hopfield-network generative method, produces synthetic patient profiles from a 23-patient pregnancy coagulation cohort and is presented as validated. FlatASCEND pairs an autoregressive sequence model with a zero-inflated log-normal time head and is tested on MIMIC-IV for pharmacological association testing. What is new across these works is the combination of irregular-time prediction, generative sampling, and downstream association testing within one sequence-modeling framework. All three items are preprints, so comparative performance across these approaches is not yet established.