Search

Transformer-based architectures have significantly advanced clinical natural language processing by improving the capture of contextual relationships in unstructured electronic health records compared to earlier recurrent and convolutional models, with domain-specific variants such as ClinicalBERT and BioBERT designed to better handle clinical terminology, abbreviations, and specialized language, thereby improving information extraction performance, although the relative impact of different pre-training strategies remains insufficiently synthesized and requires systematic evaluation of corpus selection and fine-tuning approaches; this systematic review mapped studies focusing on pre-training corpora, fine-tuning methods, and named entity recognition performance across entity types such as medications, diseases, procedures, laboratory tests, and social determinants of health, using PRISMA-guided methods and searches across PubMed, ACL Anthology, arXiv, and IEEE Xplore, identifying 32 eligible studies from 1,247 records; findings showed that ClinicalBERT, BioBERT, and PubMedBERT were the most frequently evaluated models, pre-trained on datasets such as MIMIC-III, PubMed abstracts, and mixed biomedical corpora, with consistent evidence that domain-specific pre-training outperforms general-domain BERT models on benchmarks like i2b2 and n2c2 despite variation across entity types and fine-tuning strategies, while clinical pre-training on large EHR corpora improves named entity recognition and optimized fine-tuning approaches such as lower learning rates and data augmentation further enhance performance, particularly for medications and diseases, underscoring the importance of domain adaptation and the need for more standardized evaluation protocols in clinical NLP research.

Rare diseases collectively affect over 300 million people globally, yet individual conditions are often missed due to low clinician familiarity and non-specific presenting symptoms that mimic common disorders. Supervised machine learning requires large numbers of labeled examples for training, but rare diseases have too few diagnosed cases to develop condition-specific predictive models using traditional approaches. We propose a multimodal foundation model pretrained on 10 million de-identified electronic health records (EHRs) combining clinical notes and laboratory values for zero-shot rare disease diagnosis without requiring labeled training examples. The framework comprises four components: a clinical note encoder based on a large language model, a laboratory value encoder using a time-series transformer, a multimodal fusion module with cross-attention, and a zero-shot classifier that compares patient embeddings to disease descriptions. Pretraining on large-scale EHR data enables the model to learn general medical knowledge and disease patterns, allowing diagnosis of rare conditions by recognizing manifestations even when no labeled examples of that specific disease were used for training.