MedSpaCy

Pre-trained models for identifying drugs, diseases, procedures, and lab values in unstructured text. MedSpaCy's NER components are fine-tuned on clinical corpora, handling the idiosyncratic shorthand and abbreviations common in EHR notes. It supports both span-based and token-based entity extraction, allowing for the capture of complex, multi-word clinical phrases.

Implements the ConText algorithm to determine the contextual status of clinical findings. It goes beyond simple negation to detect:

• Hypotheticality: 'If symptoms worsen, return to ED'
• Historical context: 'History of asthma'
• Family history: 'Mother had breast cancer'
• Negation: 'Patient denies chest pain' This ensures extracted entities are clinically actionable and not misleading.

A rule-based sectionizer that parses clinical notes into logical zones like Past Medical History, Medications, and Assessment/Plan. This component uses header patterns and formatting cues to segment documents, enabling downstream models to apply section-specific logic—for example, treating a medication mention in the 'Allergies' section differently than one in 'Current Medications'.

Integrates with the Unified Medical Language System (UMLS) to normalize recognized entities to unique Concept Unique Identifiers (CUIs). MedSpaCy's Linker component performs approximate nearest-neighbor search over concept embeddings, mapping ambiguous surface forms like 'MI' to the correct concept (myocardial infarction vs. mitral insufficiency) based on surrounding context.

Leverages spaCy's native component-based pipeline design. Developers can mix and match MedSpaCy's clinical components with custom rules, statistical models, or transformer backbones. The framework supports:

• Rule-based components for high-precision extraction
• Deep learning models for high-recall pattern recognition
• Hybrid approaches combining both for production robustness

A dedicated component for identifying and expanding clinical abbreviations in real-time. Using a curated dictionary of common medical shorthand, it detects tokens like 'pt' (patient), 'bid' (twice daily), and 'sob' (shortness of breath). This normalization step significantly improves downstream NER and linking accuracy by reducing lexical variability.