Clinical System One (CS1): A 4B-Parameter Decision Model for Clinical Records

Forum topic · 小凯 · 2026-10-22

Summary

Researchers introduce Clinical System One (CS1), a 4-billion-parameter decision model designed for fast, typed judgments over clinical records. CS1-4B was trained on 12.4 million question-answer pairs from 3.6 million patients across a multi-hospital health system, using structured-data keys, observed outcomes, and language-model supervision over clinical notes, diagnostic reports, and patient timelines. The model answers multiple typed clinical questions in a single forward pass without task-specific retraining, and attaches a probability to each answer so uncertain cases can be routed to human review. On 963,408 held-out patients, CS1-4B outperformed open-source decision models up to 27B parameters on 13 of 14 tasks. Across 21 external datasets, its aggregate performance approached the strongest 27B test model while running 6.4x faster under matched workloads. Integrated with language-model agents, CS1 screened a population against eligibility criteria from 17 registered clinical trials, amplifying eligible patient identification 4x over keyword search in the top 20 candidates while using 690x fewer language-model tokens. The work supports a division of labor in clinical AI: small decision models handle repetitive record reading, while generative models pose questions and synthesize evidence. arXiv: 2610.27099.

Paper Overview

Field: Machine Learning Authors: Andrew Zhang, Trinh T.L. Vuong, Sophia J. Wagner, Dongmin Bang, Tong Ding, Yuchang Su, Mary Cipperman, Christina S. Chen, Judith Lefkes, Ramin Kahidi, Ming Y. Lu, Caiwei Tian, Tianhao Luo, Jingnan Ma, Rowland Pettit, Long Phi Le, Faisal Mahmood Published: 2026-10-22 arXiv: 2610.27099

Abstract

Clinical care, health-system operations and biomedical discovery depend on extracting reliable answers from patient records, yet applying diverse clinical questions across entire populations remains computationally expensive.

The authors present Clinical System One (CS1), a 4-billion-parameter decision model for fast, typed judgments from clinical records, along with two agentic harnesses in which CS1 works together with a generalist language model.

Key details

Training data: 12.4 million question-answer pairs from 3.6 million patients across a multi-hospital health system, combining structured-data keys, observed outcomes and language-model supervision across clinical notes, diagnostic reports and patient timelines.
Inference: Answers multiple typed questions in a single forward pass without task-specific retraining; each answer carries a probability, enabling routing of uncertain cases to human review.
Benchmarks: On 963,408 held-out patients, CS1-4B exceeded open-source decision models up to 27B parameters on 13 of 14 tasks. On 21 external datasets, its aggregate performance approached the strongest 27B test model while being 6.4x faster at inference under matched workloads.
Agentic use: Integrated with language-model agents, CS1 produced cited answers covering full multi-year patient histories and screened a population against eligibility criteria from 17 registered clinical trials. In the top 20 candidates, it amplified the number of judged-eligible patients 4x versus keyword search while using 690x fewer language-model tokens.

Takeaway

The findings support a division of labor in clinical AI: small decision models perform the repetitive record reading, while generative models formulate questions and synthesize evidence.

---

*Auto-collected on 2026-10-22.*