A small clinical model screened records faster—but is it ready to guide care?
CS1-4B was trained on 12.4 million question–answer pairs from 3.6 million patients and tested on 963,408 held-out patients. Its authors say it is for research, operations and review—not autonomous patient decisions.
The short version
- What happened
- CS1-4B was trained on 12.4 million document–question pairs from 3.64 million patients across 24 Mass General Brigham sites.
- Who it affects
- A well-calibrated screening tool could help research staff find candidates sooner, but a confident error could wrongly exclude someone unless a person reviews the underlying record.
- What remains unproven
- This is an unreviewed retrospective study from one training health system. Some labels and the trial-ranking judge came from LLMs, prospective clinical outcomes were not measured and the model weights await institutional approval.
Keep in mind
This is an unreviewed retrospective study from one training health system. Some labels and the trial-ranking judge came from LLMs, prospective clinical outcomes were not measured and the model weights await institutional approval.
Source analysis · 1 linked source. Primary analysed source date: 19 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
The model reads once, then answers many narrow questions
Researchers at Mass General Brigham, Harvard and MIT report a 4.2-billion-parameter model designed to answer many predefined questions from the same clinical record in one pass. CS1-4B learned from 12,416,448 document–question pairs covering 3,643,798 patients. Labels came from structured rules, observed outcomes and GLM-5.3-Flash supervision. On 963,408 held-out patients, it led the compared open decision models on 13 of 14 tasks whose answers did not come from its training teacher.
Across 21 external datasets, CS1 scored 86.7 averaged over six task types, below the strongest 27-billion-parameter comparator at 88.6 but above its own 4B initialization at 81.7. In a matched hardware test it was 6.4 times faster than that 27B model. Its absolute outcome risks did not transfer cleanly: mean predicted risk was 0.39 where observed incidence was 0.19 before site-level recalibration.[1]
Trial screening looked promising, but clinicians did not adjudicate the headline result
The team also screened 1,000 MIMIC-IV patients against 17 registered trials. For 13 trials with at least one judged eligible candidate, the system ranked 3.5 judged-eligible patients in the top 20 on average, versus 0.8 for keyword search, while using far fewer general-model tokens. A blinded GLM-5.3 judge—not a clinical panel—graded the ranking, and the same model family wrote questions and some reference keys.
For trial teams, the approach could reduce repetitive chart reading while showing sentences that support a decision. It cannot establish eligibility or safe care on its own. The authors explicitly limit CS1 to research, operations and review with a person in the loop. Prospective evaluation, independent clinician adjudication, demographic subgroup analysis and validation at hospitals outside the training network would be needed before clinical use.[1]
The distinction that matters
How CS1 divides record screening from final judgment
A general model plans questions; the compact model scans records; a person must review uncertain or consequential cases.
Plan
Criteria become typed questions
Scope is explicit
Screen
CS1 answers across records
Probabilities and evidence sentences are returned
Review
Low-confidence or high-stakes cases escalate
Human judgment remains required