Cached Foundation Model Summaries for Memory-Efficient Clinical Time Series Inference

Accepted at the 1st Time Series in the Age of Large Models (TSALM) Workshop, ICLR 2026. [Paper]

Rafi Al Attrach, Rajna Fani, David Restrepo, Yugang Jia, Leo Anthony Celi, Peter Schüffler Massachusetts Institute of Technology | Technical University of Munich | CentraleSupelec

---

Transformer models for clinical time series face a deployment bottleneck: patient histories can span thousands of irregularly spaced events, but inference hardware imposes strict memory budgets. We study a decoupling strategy in which a pretrained foundation model compresses each patient's historical events into a fixed-size cached summary offline, and a lightweight model processes only a short recent window conditioned on that summary at inference time. Through 252 experiments on in-ICU mortality prediction (MIMIC-IV v3.1) we characterize when this is worthwhile. Cached summaries yield a 6.5% relative AUROC gain at N=8 recent events (p < 0.001), with a clear diminishing-returns pattern as N grows. FiLM modulation outperforms token injection (p < 0.001); recent-history summaries outperform distant-history summaries (p < 0.01).

---

Setup

Python 3.11 is required.

git clone https://github.com/rafiattrach/cached-summaries-ehr-inference.git
cd cached-summaries-ehr-inference
bash setup.sh
source venv/bin/activate
pip install meds-torch

Verify the installation:

python smoke_test.py

This runs three forward passes with random tensors (no MIMIC data required) and should print all three as PASS.

---

Data preparation

MIMIC-IV requires credentialed access via PhysioNet. Convert the raw dataset to MEDS format using the MIMIC-IV MEDS pipeline, then tensorize with meds-torch:

export MEDS_DIR=/path/to/mimic-iv/meds
export MODEL_DIR=/path/to/triplet_tensors
export MEDS_COHORT_DIR=/path/to/meds_cohort
export N_PARALLEL_WORKERS=8

meds-torch-ETL \
    pipeline_config_fp=experiments/configs/triplet_config.yaml \
    input_dir=$MEDS_DIR \
    output_dir=$MODEL_DIR \
    stage_runner_fp=experiments/configs/joblib_runner.yaml \
    hydra.launcher.n_jobs=$N_PARALLEL_WORKERS

---

Reproducing experiments

The environment variables below are assumed to be set for all steps. Adjust paths to match your storage layout.

export MEDS_DIR=/path/to/mimic-iv/meds
export MODEL_DIR=/path/to/triplet_tensors
export MEDS_COHORT_DIR=/path/to/meds_cohort
export SUMMARY_CACHE_DIR=$MODEL_DIR/hybrid_summary_cache
export TASK_NAME=mortality/in_icu/first_24h
export GPU=0

Step 1: extract task cohort

aces-cli --multirun hydra/launcher=joblib \
    data=sharded data.standard=meds \
    data.root="$MEDS_DIR/data" \
    "data.shard=$(expand_shards $MEDS_DIR/data)" \
    cohort_dir="$MEDS_COHORT_DIR/tasks/" \
    cohort_name="$TASK_NAME" \
    config_path="experiments/configs/tasks/mortality/in_icu/first_24h.yaml"

Step 2: precompute summary vectors (run once, reused across all training runs)

bash scripts/precompute_summaries.sh \
    --data_dir $MODEL_DIR \
    --output_dir $SUMMARY_CACHE_DIR \
    --summary_size 256 \
    --variant recent \
    --gpu $GPU

Step 3: train individual configurations

bash scripts/run_experiment.sh -m baseline -n 64 -g $GPU \
    -r $(pwd) -o $MODEL_DIR/outputs

bash scripts/run_experiment.sh -m hybrid -a film -v recent -n 64 -s 256 -g $GPU \
    -r $(pwd) -o $MODEL_DIR/outputs

bash scripts/run_experiment.sh -m oracle -n 64 -s 256 -g $GPU \
    -r $(pwd) -o $MODEL_DIR/outputs

Step 4: full grid search (reproduces all 252 runs)

bash scripts/sweep_context_budget.sh \
    --gpu $GPU \
    --epochs 10 \
    --out_dir $MODEL_DIR/outputs/sweep \
    --repo_dir $(pwd)

Ablations

bash scripts/ablate_integration_method.sh \
    --gpu $GPU --out_dir $MODEL_DIR/outputs/ablation_integration

bash scripts/ablate_summary_source.sh \
    --gpu $GPU --out_dir $MODEL_DIR/outputs/ablation_source

---

Key results

Condition	AUROC gain vs baseline	p-value
N=8, FiLM, recent	+6.5%	< 0.001
N=32, FiLM, recent	+2.1%	< 0.001
N=128, FiLM, recent	+0.4%	n.s.
N=256, FiLM, recent	+0.1%	n.s.
FiLM vs token (all N)	FiLM consistently better	< 0.001
Recent vs distant summary	Recent consistently better	< 0.01

Memory at inference scales with N only, independent of total patient history length.

---

Citation

@misc{attrach2026cached,
  title     = {Cached Foundation Model Summaries for Memory-Efficient Clinical Time Series Inference},
  author    = {Al Attrach, Rafi and Fani, Rajna and Restrepo, David and Jia, Yugang
               and Celi, Leo Anthony and Sch\"{u}ffler, Peter},
  year      = {2026},
  note      = {1st ICLR Workshop on Time Series in the Age of Large Models}
}

---

Dataset and acknowledgments

MIMIC-IV is available via PhysioNet with credentialed access. This codebase extends the meds-torch framework.