Block diffusion trades speed for accuracy on longitudinal EHR data
Original title: Block Diffusion for Longitudinal EHR: Accuracy, Efficiency, and Representation Trade-offs
On MIMIC-IV, a block size of 8 evaluates about 1.43 times faster, while linear probes nearly match autoregressive models.
Researchers compare block-based masked diffusion with autoregressive models for zero-shot prediction from longitudinal electronic health records on MIMIC-IV under matched conditions. Small blocks retain much of the autoregressive predictive performance and run faster, but generative rollout stays below autoregressive models in most settings. Frozen linear probes perform nearly identically on clinical tasks, and a larger denoising budget brings diminishing returns.
- Tests matched tokenisation, context length, model capacity, and evaluation.
- A block size of 8 evaluated about 1.43 times faster.
- Raising the denoising budget produced diminishing returns.
From the original article
1 min read at arxiv.orgAutoregressive (AR) models have proved effective for zero-shot prediction from longitudinal electronic health records, but impose strictly sequential generation on trajectories containing temporally clustered and interdependent clinical events. We investigate block-based masked diffusion (BD) as an alternative on MIMIC-IV under matched tokenisation, context length, model capacity, and evaluation conditions.
By Ashvin Gupta · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.