Conditional synthesis of training data for clinical NLP tasks in Spanish and English

Background Access to clinical narratives for NLP research is severely restricted by privacy regulations, yet these texts are essential for training models for tasks such as ICD-10 coding and clinical named entity recognition (NER). Synthetic data generated by large language models (LLMs) offers a privacy-preserving alternative, but its downstream utility under data constraints and across tasks of varying complexity remains insufficiently characterised. Methods We propose a conditional generation fram ework based on instruction-tuning a 20B-parameter LLM, where ICD-10 codes or clinical entities serve as conditioning inputs. We evaluate synthetic corpora across four datasets in English and Spanish covering discharge summaries, referral letters, and radiology reports, assessing downstream utility, corpus fidelity, and memorization-based privacy proxies. A data-scaling ablation varies the generator training fraction (5–100%), and one- and three-shot prompting of the same base model serve as baselines. Results NER models trained on synthetic data closely approach real-data performance, and even generators trained on a small fraction of real data retain most of the downstream utility. ICD-10 classification achieves comparable overall results, but per-category analysis reveals that rare diagnoses remain poorly captured. N-gram overlap stays low across all conditions, with no verbatim document reproductions foundfor fine-tuned generators. Fine-tuning outperforms prompting baselines on NER even when trained on 5% of the data, whereas for ICD-10 coding prompting is competitive. Conclusions Conditional synthesis is a viable strategy for generating clinical NLP training data, particularly for entity-level tasks. Task complexity and label imbalance, rather than data quantity, emerge as the primary determinants of the remaining performance gap.

Authors

Publication Details

Journal
BMC Medical Informatics and Decision Making
Published
2026-11-19
DOI
https://doi.org/10.1186/s12911-026-03962-8
Primary Topic
Machine Learning in Healthcare
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Conditional synthesis of training data for clinical NLP tasks in Spanish and English

Jocelyn Dunstan, Thomas Vakili, Claudio Aracena, Fabián Villena et al.
BMC Medical Informatics and Decision Making
Machine Learning in Healthcare
article

Conditional synthesis of training data for clinical NLP tasks in Spanish and English

Jocelyn Dunstan, Thomas Vakili, Claudio Aracena, Fabián Villena, Luis Miranda
article en

Abstract

Background Access to clinical narratives for NLP research is severely restricted by privacy regulations, yet these texts are essential for training models for tasks such as ICD-10 coding and clinical named entity recognition (NER). Synthetic data generated by large language models (LLMs) offers a privacy-preserving alternative, but its downstream utility under data constraints and across tasks of varying complexity remains insufficiently characterised. Methods We propose a conditional generation fram ework based on instruction-tuning a 20B-parameter LLM, where ICD-10 codes or clinical entities serve as conditioning inputs. We evaluate synthetic corpora across four datasets in English and Spanish covering discharge summaries, referral letters, and radiology reports, assessing downstream utility, corpus fidelity, and memorization-based privacy proxies. A data-scaling ablation varies the generator training fraction (5–100%), and one- and three-shot prompting of the same base model serve as baselines. Results NER models trained on synthetic data closely approach real-data performance, and even generators trained on a small fraction of real data retain most of the downstream utility. ICD-10 classification achieves comparable overall results, but per-category analysis reveals that rare diagnoses remain poorly captured. N-gram overlap stays low across all conditions, with no verbatim document reproductions foundfor fine-tuned generators. Fine-tuning outperforms prompting baselines on NER even when trained on 5% of the data, whereas for ICD-10 coding prompting is competitive. Conclusions Conditional synthesis is a viable strategy for generating clinical NLP training data, particularly for entity-level tasks. Task complexity and label imbalance, rather than data quantity, emerge as the primary determinants of the remaining performance gap.

BMC Medical Informatics and Decision Making
Machine Learning in Healthcare
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon