On-Premises Small Language Models Accurately Classify Pregnancy Status in Obstetric Reports
Researchers have evaluated seven small language models running on local servers to classify pregnancy status from clinical ultrasound reports while maintaining data privacy.
Key points
- Evaluated seven small language models for pregnancy status classification
- Achieved high macro F1 scores on both MIMIC-IV and international OCR cohorts
- Demonstrated privacy-preserving, on-premises data processing without fine-tuning
A recent study published in Nature Health Sciences investigated the use of compact, on-premises language models to extract critical pregnancy status information from unstructured obstetric ultrasound texts. Led by researchers from institutions including Saint Mary's University and York University, the evaluation tested seven frozen models with 3 to 4 billion parameters on 628 reports from the MIMIC-IV database and 340 optical character recognition (OCR) reports from an Indian clinical site. The systems categorized texts into early active pregnancy, late active pregnancy, or no active pregnancy based on a 14-week operational boundary.
Study Performance and Methodology
Using 12 in-context exemplars, a committee of seven models attained a macro F1 score of 0.9622 on the MIMIC-IV cohort and 0.9814 on the Indian OCR group. The investigators also introduced a probability-guided exemplar selection method to minimize annotation workloads. This approach successfully converted 95% to 98% of annotation effort into usable balanced prompt examples, compared to 52% to 66% achieved through standard random sampling techniques.
- What was studied: Seven small language models analyzing obstetric ultrasound reports.
- Key finding: High classification accuracy achieved locally without requiring offsite data transfer or extensive fine-tuning.
- Study limitations: Retrospective analysis relying on pre-existing, anonymized datasets and clinical texts.
What to know: The retrospective findings suggest that local artificial intelligence tools may assist clinical informatics in processing sensitive healthcare data securely, though human review remains essential. Always consult qualified healthcare professionals for medical guidance and personal health decisions.
Source Nature Health Sciences — nature.com/articles/s41598-026-79573-0 ↗
Read next
❤️Conditions & CareMigraine and Smoking Linked to Increased Early Stroke Risk in Women
A large-scale study suggests that people who have given birth and experience migraines while also smoking face a significantly higher risk of stroke before age 60.
🔬ResearchCBD Oil Found Ineffective for Managing Behavioral Issues in Youths with Intellectual Disabilities
A double-blind clinical trial published in JAMA Pediatrics indicates that CBD oil provides no significant behavioral improvement for children and teenagers with intellect…
🏃FitnessExpert Advice on Hydration Gear Sparks Debate Among Recreational Runners
Olympic medalist Josh Kerr recently suggested that runners may not require hydration packs for distances under 12 miles, igniting a broader conversation about gear necess…
🏃FitnessA Structured Approach to Beginning a Strength and Cardio Routine
A new four-week training program offers beginners a guided path to incorporating strength training and cardiovascular exercise into their daily lives.
🏥Public HealthRecall Issued for Generic Propranolol Over Potential Metal Contamination
Breckenridge Pharmaceutical has initiated a voluntary recall of one lot of generic propranolol capsules due to the potential presence of metal fragments.
🥗NutritionUnderstanding Food Science: Why Some Leftovers Freeze Better Than Others
Freezing leftovers is an effective way to minimize food waste, though the success of the process depends on the cellular structure and composition of the food being store…