← News

Models & toolsSources: arXiv · 18 Nov 2026

Multimodal LLM system extracts scores from scanned cognitive assessment forms

A modular pipeline with human oversight reads handwritten clinical forms, reporting large time and cost savings over manual data entry.

In plain words

Clinics often record cognitive tests on paper, and typing those results into computers is slow and error-prone. Researchers built a system that classifies each scanned page and uses multimodal large language models to pull out fields and scores, with people reviewing only the hardest pages.

What was announced

The system, described in an arXiv paper, works in two stages: a form classifier assigns each scanned page to a type such as MMSE, MoCA, Mini-Cog or the NICHQ Vanderbilt Assessment, then a matching score extractor pulls fields and scores into a schema. A dual-model validation step brings in human annotators only when two models disagree, which the authors say limits manual review to difficult pages. Reported end-to-end accuracy was 88% for cognitive assessments, rising to 97.5% with the human-in-the-loop step, while the NICHQ Vanderbilt reached 97.5% overall. The authors also state the approach cut manual work by 63.8% and, versus manual extraction, reduces processing time by 10 to 20 times and costs by more than 12 times.

Accuracy varied by task. In the Mini-Cog Clock Drawing analysis, the authors found 62.3% of extraction errors came from ambiguous hand-drawn clocks, and in those cases the model's readings were described as reasonable. The modular design, they argue, allows new form types to be added without rebuilding the pipeline.

What it does not show

  • All figures come from the paper's own abstract; no independent evaluation or deployment results are available.
  • The source is an abstract only, so methods detail, dataset size and error definitions cannot be verified.

For security and IT teams

Scope
The reported system targets scanned cognitive and behavioral assessment forms (MMSE, MoCA, Mini-Cog, NICHQ Vanderbilt) rather than general document extraction.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.