Skip to main content

Entity Extraction3.0.5

CCP Product Release Notes

About

This release notes page highlights changes, improvements, and updates available for a new version release of the Clinical Comprehensive Pipeline (CCP).

The Clinical Comprehensive Pipeline (CCP) is an NLP pipeline that recognizes and extracts clinical entities and modifiers from unstructured text across IMO’s problem/diagnosis, procedure, medication, laboratory, and clinical observation domains. Entities extracted from unstructured text are also mapped to the following standard terminologies and code systems:

  • Problem/Diagnosis domain: ICD-10-CM, ICD-9-CM, SNOMED CT, and IMO lexicals.
  • Procedure domain: CPT (Current Procedural Terminology), HCPCS (Healthcare Common Procedure Coding System), ICD-10-PCS, SNOMED CT, LOINC (Logical Observation Identifiers Names and Codes), and IMO lexicals.
  • Medication domain: RxNorm, NDC (National Drug Code), CVX, IMO lexicals.
  • Lab Domain: LOINC, CPT, HCPCS, ICD-10-PCS, SNOMED CT, and IMO lexicals.

CCP Version 3.0.3

What’s New in Version 3.0.3

Release Date: 05/04/2026

CCP v3.0.3 delivers significant accuracy gains across all core entity types: problem, procedure, lab, medication, clinical observation, and social factors.

These improvements are driven by two areas:

  • Model Development: An expanded and diversified training data corpus, which grew from 2000 notes to 7000 notes (317% growth in total entity volume).
  • Model Evaluation: A more rigorous evaluation methodology, backed by an expert-created Gold Standard dataset that increased by 111% in total entity volume.

Model Development: Training Data Expansion

Key aspects of the model development effort include:

  • Larger training set: 5,000 additional expert-annotated notes provide the models with significantly more examples of clinical language variation, improving generalization to unseen text.
  • Increased entity diversity: The expanded dataset more than doubled the annotated entity count across all semantic types, with social factor entities seeing the largest relative growth (+179%), addressing a previously underrepresented category.
  • Broader note type coverage: Training on a wider variety of clinical note types reduces bias toward any single documentation style.

Training Corpus Entity Volume Growth

The training corpus grew from 89,996 entities in v3.0 to 375,339 entities in v3.0.3, which is a 316% increase in total entity volume across all semantic types. Problem entities represent the largest category (45%) of the training corpus.

Entity Type Old Training Entities (v3.0) New Training Entities (v3.0.3) Growth
Problem 39,749 170,289 +328%
Procedure 9,397 38,375 +308%
Lab 12,332 43,464 +252%
Medication 8,048 34,316 +326%
Clinical Observation 8,929 35,068 +293%
Social Factor 1,500 7,330 +389%
Total 89,996 375,339 +316%

Model Evaluation: Gold Standard Expansion

Accuracy was evaluated against a Gold Standard dataset, which was created and validated by a team of IMO Clinical Subject Matter Experts. The Gold Standard was designed to reflect a diverse range of real-world clinical documentation, and included the characteristics below:

  • Note types: Outpatient Visit, Progress, History & Physical (H&P), Consultation, Case Reports, ED, Discharge Summaries, Procedure Reports, SOAP notes, and Operative Notes.
  • Formats: Typed, dictated, and transcribed electronic text notes.
  • Clinical settings: Data was sampled from both inpatient and outpatient emergency department settings.
  • Time frame: Data was sampled from 2022–2024.
  • Geographic diversity: Data captured across both rural and suburban care settings.

Gold Standard Diversification and Entity Volume Growth

The new IMO Gold Standard increases clinical coverage, more than doubling the size (+111%) of our evaluation dataset.

Semantic Type Old Gold Entities (v3.0) New Gold Entities (v3.0.3) Growth
Problem 1,471 3,233 +120%
Procedure 452 916 +103%
Lab 509 843 +66%
Medication 384 954 +148%
Clinical Observation 218 441 +102%
Social Factor 38 106 +179%
Total 3,072 6,493 +111%

Accuracy Improvements in Version 3.0.3

At the entity extraction level, accuracy for recognizing and extracting entities improved by the following increments. Note that accuracy is defined as F1 scores, and the improvement is quantified by percentage increase, not absolute increase.

  • Problem: accuracy improved by +7.5% compared to the prior version
  • Procedure: accuracy improved by +5.5% compared to the prior version
  • Medication: accuracy improved by +3.4% compared to the prior version
  • Lab: accuracy improved by +8.0% compared to the prior version
  • Clinical Observation: accuracy improved by +9.7% compared to the prior version
  • Social Factor: accuracy improved by +14.0% compared to the prior version

At the entity resolution level, accuracy for resolving extracted entities to IMO lexicals improved by the following increments. Note that accuracy is defined as F1 scores, and the improvement is quantified by percentage increase, not absolute increase.

  • Problem: accuracy remained stable with a marginal improvement compared to the prior version
  • Procedure: accuracy degraded by -6.2%; however, this is more reflective of the robust Gold Standard and diverse training dataset. The pipeline now extracts a more precise set of procedure entities, which are evaluated against a larger sample size of procedure entities in our Gold Standard.
  • Medication: accuracy improved by +18.8% compared to the prior version.
  • Lab: accuracy improved by +2.9% compared to the prior version
  • Clinical Observation: accuracy improved by +13.7% compared to the prior version

How Was Accuracy Measured?

Accuracy was evaluated against a Gold Standard dataset, which was created and validated by a team of IMO Clinical Subject Matter Experts. The Gold Standard was designed to reflect a diverse range of real-world clinical documentation, and included the characteristics below:

  • Note types: History & Physical (H&P) notes, Progress Notes, Discharge Summaries, Procedure Reports, SOAP notes, and Operative Notes.
  • Formats: Typed, dictated, and transcribed electronic text notes.
  • Clinical settings: Data was sampled from both inpatient and outpatient emergency department settings.
  • Time frame: Data was sampled from years 2022–2024.
  • Geographic diversity: Data captured across both rural and suburban care settings.

Our clinical team of physicians, informaticists, and medical coders:

  • Reviewed and annotated the Gold Standard dataset to ensure consistent application of entity and modifier definitions.
  • Validated extraction results against established IMO lexicals, which resolve to terminology and codes such as ICD-10-CM, CPT, SNOMED CT, RxNorm,etc.
  • Conducted inter-annotator agreement checks to ensure consistency in clinical interpretation.
  • Iteratively refined annotation guidelines for training, to ensure coverage of edge cases.

Release Notes

Date Version Release Summary
05/04/2026 3.0.3 Significant accuracy gains across all core entity types driven by expanded training data (2000 to 7000 notes, +316% entity volume) and a more rigorous Gold Standard evaluation dataset (+111% entity volume).

IMPROVEMENT
  • Entity extraction accuracy improved across all semantic types: Problem (+7.5%), Procedure (+5.5%), Lab (+8.0%), Medication (+3.4%), Clinical Observation (+9.7%), Social Factor (+14.0%).
  • Entity resolution accuracy improved for Medication (+18.8%), Clinical Observation (+13.7%), and Lab (+2.9%).
  • Training corpus expanded from 89,996 to 375,339 annotated entities.
  • Gold Standard evaluation dataset expanded from 3,072 to 6,493 entities with broader note type coverage.
  • Pre-processing enhancements introduced to normalize note formatting reducing input noise and improving consistency for downstream entity detection.
08/18/2025 3.0.0 Significant accuracy improvements in entity extraction and modifier detection across the pipeline.

FEATURE
  • Alignment with IMO’s clinical domains and external code systems. In the entity type schema, tests are changed to labs, drugs are changed to medications, and treatments are changed to procedures.
  • New entity types: social factors and clinical observations.
  • New modifiers for procedures and labs that provide higher specificity when resolving to standard codes.

IMPROVEMENT
  • Improved precision in procedure and lab entity identification.
  • Improved handling of historical clinical information to distinguish between active and past events.
  • Improved handling of allergies to medications.
  • Enhanced section identification for more accurate segmentation of clinical note text.
09/13/2023 2.0.0 Platform-level enhancements for security, performance and reliability.
1.0.0 Initial release

 

Notices

©️ 2026 Intelligent Medical Objects, Inc. All Rights Reserved.

CPT®️ copyright 2025 American Medical Association. All rights reserved.

SNOMED®️ and SNOMED CT®️ are registered trademarks of IHTSDO.

LOINC®️ is a registered United States trademark of Regenstrief Institute, Inc.

RxNorm is publicly available data courtesy of the U.S. National Library of Medicine (NLM), National Institutes of Health, Department of Health and Human Services.