Cached Summary Embeddings for Memory-Efficient EHR Inference

Rafi Al Attrach, Rajna Fani, David Restrepo, Yugang Jia, Leo Anthony Celi, Peter Schuffler
Proceedings of the 7th Conference on Health, Inference, and Learning, PMLR 333:778-793, 2026.

Abstract

Transformer-based clinical prediction models face a deployment challenge: processing long patient histories can require memory that exceeds available resources in resource-constrained settings. We propose a deployment architecture that separates expensive historical encoding from lightweight inference. In an offline preprocessing phase, a clinical language model compresses each patient’s historical events into a fixed-size vector (768 dimensions, 5 KB per patient). At inference time, the prediction model processes only a short window of recent events, conditioned on the cached summary. Through 252 experiments on a 24-hour in-ICU mortality cohort from MIMIC-IV, we characterize when this architecture provides value. The benefit of cached summaries decays as the recent context window grows: a 6.5% relative AUROC improvement at $N$=8 recent events ($p < 0.001$) shrinks to a negligible 0.1% at $N$=256 (not statistically significant). We find that Feature-wise Linear Modulation (FiLM) outperforms token injection for integrating summaries ($p < 0.001$). Our results provide deployment guidance: when hardware constraints limit the recent context to 32 events or fewer, cached summaries recover meaningful predictive signal; when longer sequences are feasible, the caching overhead is not justified.

Cite this Paper

BibTeX
@InProceedings{pmlr-v333-al-attrach26a, title = {Cached Summary Embeddings for Memory-Efficient EHR Inference}, author = {Al Attrach, Rafi and Fani, Rajna and Restrepo, David and Jia, Yugang and Celi, Leo Anthony and Schuffler, Peter}, booktitle = {Proceedings of the 7th Conference on Health, Inference, and Learning}, pages = {778--793}, year = {2026}, editor = {Healey, Elizabeth and Fries, Jason and Pollard, Tom and Tang, Shengpu and Zink, Anna and Hartvigsen, Tom and Agrawal, Monica and Finlayson, Sam and Glicksberg, Benjamin and Beaulieu-Jones, Brett and Wang, Kai and Fontalvo, Daseyra and Sarker, Tasmie and Chen, Irene and Alsentzer, Emily}, volume = {333}, series = {Proceedings of Machine Learning Research}, month = {29--30 Jun}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v333/main/assets/al-attrach26a/al-attrach26a.pdf}, url = {https://proceedings.mlr.press/v333/al-attrach26a.html}, abstract = {Transformer-based clinical prediction models face a deployment challenge: processing long patient histories can require memory that exceeds available resources in resource-constrained settings. We propose a deployment architecture that separates expensive historical encoding from lightweight inference. In an offline preprocessing phase, a clinical language model compresses each patient’s historical events into a fixed-size vector (768 dimensions, 5 KB per patient). At inference time, the prediction model processes only a short window of recent events, conditioned on the cached summary. Through 252 experiments on a 24-hour in-ICU mortality cohort from MIMIC-IV, we characterize when this architecture provides value. The benefit of cached summaries decays as the recent context window grows: a 6.5% relative AUROC improvement at $N$=8 recent events ($p < 0.001$) shrinks to a negligible 0.1% at $N$=256 (not statistically significant). We find that Feature-wise Linear Modulation (FiLM) outperforms token injection for integrating summaries ($p < 0.001$). Our results provide deployment guidance: when hardware constraints limit the recent context to 32 events or fewer, cached summaries recover meaningful predictive signal; when longer sequences are feasible, the caching overhead is not justified.} }
Endnote
%0 Conference Paper %T Cached Summary Embeddings for Memory-Efficient EHR Inference %A Rafi Al Attrach %A Rajna Fani %A David Restrepo %A Yugang Jia %A Leo Anthony Celi %A Peter Schuffler %B Proceedings of the 7th Conference on Health, Inference, and Learning %C Proceedings of Machine Learning Research %D 2026 %E Elizabeth Healey %E Jason Fries %E Tom Pollard %E Shengpu Tang %E Anna Zink %E Tom Hartvigsen %E Monica Agrawal %E Sam Finlayson %E Benjamin Glicksberg %E Brett Beaulieu-Jones %E Kai Wang %E Daseyra Fontalvo %E Tasmie Sarker %E Irene Chen %E Emily Alsentzer %F pmlr-v333-al-attrach26a %I PMLR %P 778--793 %U https://proceedings.mlr.press/v333/al-attrach26a.html %V 333 %X Transformer-based clinical prediction models face a deployment challenge: processing long patient histories can require memory that exceeds available resources in resource-constrained settings. We propose a deployment architecture that separates expensive historical encoding from lightweight inference. In an offline preprocessing phase, a clinical language model compresses each patient’s historical events into a fixed-size vector (768 dimensions, 5 KB per patient). At inference time, the prediction model processes only a short window of recent events, conditioned on the cached summary. Through 252 experiments on a 24-hour in-ICU mortality cohort from MIMIC-IV, we characterize when this architecture provides value. The benefit of cached summaries decays as the recent context window grows: a 6.5% relative AUROC improvement at $N$=8 recent events ($p < 0.001$) shrinks to a negligible 0.1% at $N$=256 (not statistically significant). We find that Feature-wise Linear Modulation (FiLM) outperforms token injection for integrating summaries ($p < 0.001$). Our results provide deployment guidance: when hardware constraints limit the recent context to 32 events or fewer, cached summaries recover meaningful predictive signal; when longer sequences are feasible, the caching overhead is not justified.
APA
Al Attrach, R., Fani, R., Restrepo, D., Jia, Y., Celi, L.A. & Schuffler, P.. (2026). Cached Summary Embeddings for Memory-Efficient EHR Inference. Proceedings of the 7th Conference on Health, Inference, and Learning, in Proceedings of Machine Learning Research 333:778-793 Available from https://proceedings.mlr.press/v333/al-attrach26a.html.

Related Material