MoRGen: Mixture-of-Resolutions Generative Forecasting for Irregularly Sampled Medical Time-Series Data

Autoregressive generative models for irregularly sampled clinical time-series data are increasingly used for zero-shot risk forecasting. Prior work typically adopts a single fine-grained discretization of time, where tokens are generated at one fixed, predetermined temporal resolution. We demonstrate that the zero-shot accuracy of individual generative forecasters varies with temporal resolution: performance can degrade when the model resolution is poorly matched to the temporal dynamics of the endpoint being evaluated. We then propose MoRGen (Mixture-of-Resolutions Generation), which fuses forecasts from generative experts trained at multiple temporal resolutions using a low-capacity task-specific mixture, improving performance across tasks with different temporal dynamics. Across multiple horizons and outcomes on three independent clinical datasets, MoRGen achieves lower binary cross-entropy (BCE) and statistically significant AUROC gains over autoregressive generative models that forecast tokens at a fixed temporal resolution.

Doctors often need to estimate a patient’s risk of future health events, such as hospital readmission, death, or abnormal lab results, to guide follow-up care after discharge. Recent machine learning models can simulate possible future medical records, but they usually generate these futures at one fixed time scale, such as minute by minute. This can be inefficient and unreliable: some clinical changes happen quickly, while others unfold over weeks, months, or years. We developed MoRGen, a method that combines several generative forecasting models, each trained to look at the future using a different time resolution. Instead of assuming one time scale is best for every medical outcome, MoRGen learns how to combine forecasts from fine and coarse time scales for each prediction task. Across three clinical datasets and multiple outcomes, MoRGen produced more accurate risk estimates than a strong single-resolution generative baseline, especially for longer-term outcomes such as mortality. Our results suggest that choosing the right time scale is an important part of building reliable clinical forecasting models. More broadly, combining forecasts across time scales may improve the accuracy of generative medical AI for post-discharge risk prediction.