Harvard's CS1 Reads Patient Records 6.4x Faster Than 27B Models

Screening years of patient records against dozens of criteria requires repeated passes through clinical notes, lab timelines, discharge summaries, and diagnostic reports. General-purpose large language models can process those records, but token use and inference costs climb quickly across thousands of patients. Researchers at Mass General Brigham and Harvard propose dividing the work: a small decision model handles high-volume record review, while a generalist LLM plans queries and synthesizes the results.

Clinical System One, or CS1, is a 4-billion-parameter model trained to return fast, typed judgments from clinical records. The researchers trained CS1-4B on 12.4 million question-answer pairs derived from 3.6 million patients in a multi-hospital health system. The data combines structured fields, observed outcomes, and language-model supervision across clinical notes, diagnostic reports, and longitudinal timelines.

The name refers to Daniel Kahneman’s dual-process framework. CS1 serves as the fast pattern-matching component, while a general-purpose LLM performs slower planning and synthesis.

Typed answers cut the workload

Many record-review tasks can be expressed as constrained questions: Does this note mention metformin? Was the patient discharged on an anticoagulant? Is creatinine rising, falling, or stable? CS1 selects from predefined response options and assigns each option a probability instead of generating free-form prose.

A single forward pass can answer multiple questions about one document, without task-specific retraining for each criterion. That interface lets applications submit batches of structured questions, apply probability thresholds, and send uncertain cases to human reviewers.