Research map / Large language models / AI and society
ACES: Automatic Cohort Extraction System for Event-Stream Datasets
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces ACES, a pip-installable configuration language and recursive extraction system that pulls labeled patient cohorts from event-stream datasets (MEDS, ESGPT, or direct predicates) so task definitions can be shared and reproduced conceptually across datasets.
Problem
Reproducing ML-for-healthcare cohort and task definitions is hard because cohorts are often under-specified and data is private; existing cohort tools (ATLAS, i2b2 PIC-SURE, OMOP-learn, DPM360) are tied to specific common data models or are hard to integrate with deep learning workflows.
Approach
Separate dataset-specific event predicates from dataset-agnostic, interrelated temporal/event-bounded windows in a shareable configuration file, and execute them with a recursive tree-based algorithm over event-stream data, exposed via CLI and Python API.
Key results
- On a single MIMIC-IV MEDS shard (47,954 patients, 80.3M rows), the nine example tasks ran in roughly 183-368 s total time, with up to ~109 GB peak memory for the mortality and readmission tasks.
- In a preliminary comparison on Synthea-1000, ACES via MEDS ran the first-24h in-hospital mortality task in 0.386 s with 35 lines of code, versus 5.932 s and 205 lines for DPM360 via OMOP.
- ACES required 22-35 lines of code across four tasks in the comparison, versus 172-288 lines for OMOP-learn and DPM360, and was roughly 10-60x faster in runtime on the reported tasks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ACES via MEDS instead of DPM360 via OMOP | Cohort extraction · Synthea-1000, First 24h in-hospital mortality | Runtime (s): 5.932 -> 0.386 | Helped |
| Use ACES via MEDS instead of DPM360 via OMOP | Cohort extraction · Synthea-1000, 30d post-hospital-discharge mortality | Runtime (s): 4.188 -> 0.236 | Helped |
| Use ACES via MEDS instead of DPM360 via OMOP | Cohort extraction · Synthea-1000, 30d re-admission | Runtime (s): 6.260 -> 0.337 | Helped |
| Use ACES via MEDS instead of DPM360 via OMOP | Cohort extraction · Synthea-1000, End-of-life prediction | Runtime (s): 4.901 -> 0.449 | Helped |
| Use ACES via MEDS instead of OMOP-learn via OMOP | Cohort extraction · MIMIC-IV Demo, First 24h in-hospital mortality | Runtime (s): 12.220 -> 0.617 | Helped |
| Use ACES via MEDS instead of OMOP-learn via OMOP | Cohort extraction · MIMIC-IV Demo, 30d post-hospital-discharge mortality | Runtime (s): 8.608 -> 0.301 | Helped |
| Use ACES via MEDS instead of OMOP-learn via OMOP | Cohort extraction · MIMIC-IV Demo, 30d re-admission | Runtime (s): 19.710 -> 0.455 | Helped |
| Use ACES via MEDS instead of OMOP-learn via OMOP | Cohort extraction · MIMIC-IV Demo, End-of-life prediction | Runtime (s): 24.540 -> 0.349 | Helped |
Limitations
- Predicates over unstructured data (notes, images) must currently be extracted manually by users.
- Comparative experiments are preliminary, on small datasets, and the authors acknowledge potential bias from their familiarity with ACES.
- Direct interoperability with ATLAS and OHDSI vocabulary-derived cohort definitions is not yet supported.
Experimental setup
Datasets: MIMIC-IV (MEDS version) MIMIC-IV OMOP Demo Synthea-1000 (synthetic OMOP)
Benchmarks: MIMIC-IV MEDS shard computational profile Synthea-1000 comparison MIMIC-IV Demo comparison
Metrics: runtime (s) peak memory (MiB) lines of code human time (s) # patients # samples
Baselines: OMOP-learn DPM360 OHDSI ATLAS
Builds on: MEDS Event Stream GPT Hydra
Closest papers by meaning
- NeuralCohort: Cohort-aware Neural Representation Learning for Healthcare Analytics (ICML 2025 · similarity 0.88)
- One Loss to Rule Them All: Marked Time-to-Event for Structured EHR Foundation Models (NeurIPS 2026 · similarity 0.84)
- Exploiting Negative Samples: A Catalyst for Cohort Discovery in Healthcare Analytics (ICML 2024 · similarity 0.84)
- Context Clues: Evaluating Long Context Models for Clinical Prediction Tasks on EHR Data (ICLR 2025 · similarity 0.84)
- Contrastive Learning for Clinical Outcome Prediction with Partial Data Sources (ICML 2024 · similarity 0.83)
- Automated Multi-Task Learning for Joint Disease Prediction on Electronic Health Records (NeurIPS 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.