Abstract
Electronic health records (EHRs) contain dense, longitudinal free-text data that remain inaccessible to computational analysis. In orthodontics, these progress notes encode complex biomechanical and compliance information using non-standard shorthand, representing a prototypical structured extraction challenge. Converting them to structured data is essential for large-scale analysis, but remains unsolved. Manual chart review remains the research standard, bottlenecked by time-consuming note-by-note evaluation. Commercial cloud LLMs impose API costs and HIPAA constraints, struggle with specialized shorthand, and fail strict formatting, while no locally deployable method has been validated here. We present a computationally efficient, privacy-preserving workflow that fine-tunes a local 7/8B parameter open-weight LLM with LoRA to map noisy orthodontic shorthand to 61 Common Data Elements. We benchmark candidate architectures, comparing Qwen 2.5-7B and Meta-Llama-3.1-8B, and evaluate the selected model against two references on matched note sets: a frontier model (Gemini 3 Pro) and a human inter-rater ceiling. From 6,419 notes, the fine-tuned model reached a weighted F1 of 0.740 on the test set, outperformed the frontier baseline on matched notes by 53%, and achieved 80% of the human ceiling while executing 57 times faster than manual review. Crucially, it remains deployable as a 4.66 GB quantized model on standard clinician hardware. This work establishes a replicable, privacy-by-design infrastructure for clinical NLP in specialized domains, with quantized model weights made publicly available. While downstream clinical analyses fall outside this scope, this cost-effective method establishes the essential infrastructure for future orthodontic research.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/EMA_013.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: Not Available
BibTex
@InProceedings{DumPau_Efficient_MICCAISAT2026,
author = { Dumont, Paul AND Buisson, Alexandre AND Teixeira, Rodrigo AND Aliaga, Aron AND Yatabe, Marilia AND Li, Tengfei AND Dai, Runpeng AND Zhu, Hongtu AND Caleme, Eduardo AND Barone, Selene AND Miranda, Felicia AND Mattos, Claudia Trindade AND Kuo, I-Hsien AND de Oliveira, Pedro José AND Bianchi, Jonas AND Cevidanes, Lucia H.},
title = { { Efficient Fine-Tuning of Locally Deployable LLMs for Structured Data Extraction from Longitudinal Orthodontic Progress Notes } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17265},
month = {pending},
page = {pending}
}