CAE Product Release Notes

Version: 1.0.0

Release Date: June 2026

What's Included in CAE v1.0.0

Context-Aware Extraction (CAE) is an API that analyzes full unstructured clinical notes to return the most clinically appropriate, billing-specific IMO lexical outputs for clinical problems. Unlike sentence-level extraction, which treats each sentence independently, CAE interprets clinical evidence across multiple sentences and sections of a note to resolve problems to the most specific downstream ICD-billable coded output. This document-level understanding reduces manual reconciliation for coders and delivers higher-fidelity, reimbursement-relevant structured data for downstream workflows.

Note: For the initial CAE v1.0.0 release, Context-Aware Extraction is focused exclusively on the Problem domain.

Key Capabilities

• Full-note input: Accepts clinical notes such as progress notes, discharge summaries, H&Ps, admission notes, and procedure reports.
• Billing-specific outputs: Returns the most specific IMO lexical outputs and ICD codes, grounded in source-note evidence.
• Explainability: Each output includes the evidence spans supporting why it was returned.

Model Development

CAE leverages a novel hybrid architecture combining frontier Large Language Models with an ensemble of advanced clinical NLP techniques — including neural entity recognition, vector-based semantic matching, and multi-pass contextual reasoning — to achieve deep, holistic understanding of clinical documentation that traditional single-model approaches cannot match. Unlike sentence-level extraction pipelines, CAE evaluates information across multiple sections and sentences to identify relevant clinical context, improve specificity, and support more accurate normalization to IMO terminology. By incorporating document-level context, CAE can identify clinically relevant details distributed throughout a note and generate more specific, clinically meaningful outputs.

Model Evaluation Results

CAE v1.0.0 was evaluated against the 100-note gold standard dataset. Accuracy is reported using F1 score, which balances precision and recall across extracted clinical problem concepts.

Release Notes