Answers are grounded in this paper and cite their sources.
Ultrasound reports in obstetric care provide information about the status of the pregnancy which is essential for clinical informatics purposes. This information, however, is buried in free text. We evaluated seven frozen small language models with 3–4 billion parameters, run on-premises, on 628 MIMIC-IV reports and 340 reports processed with OCR from one Indian clinical site. The models were used to classify the reports as early active pregnancy, late active pregnancy, or no active pregnancy with a 14-week operational boundary and a labeling scheme based on management status. With 12 in-context exemplars, a seven-model committee achieved a macro F1 of 0.9622 on MIMIC-IV and 0.9814 on the Indian OCR cohort. When MIMIC-IV exemplars were transferred to the Indian OCR cohort, performance did not differ detectably from that obtained with Indian OCR cohort exemplars in a paired bootstrap test. We also investigated how to reduce labeling burden and introduced a probability-guided exemplar selection procedure. The procedure converted 95–98% of annotation effort into usable balanced prompt exemplars, compared with 52–66% under random sampling. These retrospective results support pregnancy-status extraction with human review in low-annotation settings where data remain on-premises, without fine-tuning or offsite data transfer.
No takes yet. Share an insight, caveat, or question.
Neveditsin et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: