Re-Work DL Summit - Can AI Understand Doctor's Notes
1 like216 views
Change Healthcare is developing AI to understand unstructured medical text like doctors' notes to improve medical coding, reduce errors and costs. Their AI models process over 50,000 PDF pages daily and can identify diagnoses, codes, and other key information with over 90% accuracy. Change Healthcare is also using AI to enhance claims processing, clinical workflows, and patient engagement. Developing trust in AI solutions and addressing healthcare's complex data and workflows remain ongoing challenges.
Re-Work DL Summit - Can AI Understand Doctor's Notes
- 1. © 2019 Change Healthcare LLC and/or one of its subsidiaries. All Rights Reserved.© 2019 Change Healthcare LLC and/or one of its subsidiaries. All Rights Reserved. Alex Ermolaev, Director of AI 05.24.2019 CAN AI UNDERSTAND DOCTOR’S NOTES?
- 2. © 2018 Change Healthcare LLC and/or its subsidiaries or affiliates. All Rights Reserved. • Extensive product/services, expertise, data, and connectivity • Tying our success to our customer: providers, payers and consumers • Independent, unbiased, & relentless drive towards better healthcare • Delivering practical innovation through technology and partners 2,100 Payer Connections 1 in 5 US Patient Records $1 trillion Healthcare Claims Nearly 14 billion Healthcare Transactions 5,500 Hospitals & Health Systems 900,000 Physicians117,000 Dentists 600 Laboratoriesh Pervasive across healthcare The Change Healthcare Intelligent Healthcare NetworkTM
- 3. © 2019 Change Healthcare LLC and/or one of its subsidiaries. All Rights Reserved. 3 Change Healthcare initial AI focus areas With much more to come Claims Lifecycle Management helping payers and providers create efficiencies and reduce the cost burden in claims processing. Medical Text Understanding helping to improve and streamline medical coding, reduce errors, fraud, waste and abuse. Workflow Prioritization and Optimization helping make our own and our customer’s work smarter and more efficient. Patient/Member Engagement helping to increase consumer/patient experience and engagement by pairing AI with behavioral science.
- 4. © 2018 Change Healthcare LLC and/or its subsidiaries or affiliates. All Rights Reserved. • Process around 50K PDF pages per day, which is almost 20 millions over the last year • The documents are often in bad shape: badly scanned, typos, incomplete, mishandled • Finding relevant information (including codes) to make decisions on behalf of our insurance partners Medical text understanding at scale
- 5. © 2018 Change Healthcare LLC and/or its subsidiaries or affiliates. All Rights Reserved. Example
- 6. © 2018 Change Healthcare LLC and/or its subsidiaries or affiliates. All Rights Reserved. Solution Blocks ID/Class Models: TensorFlow Keras Output Reorder R to L B to T
- 7. © 2018 Change Healthcare LLC and/or its subsidiaries or affiliates. All Rights Reserved. Medical Record Pre-Processing Header ✓ Footer ✓ David A. Smith, M.D. ✓ Legalese ✓ MR Section ✓ Likely diagnoses ✓ Member info ✓
- 8. © 2018 Change Healthcare LLC and/or its subsidiaries or affiliates. All Rights Reserved. We found the relevant pieces of text. Throw everything else away. Medical Record Pre-processing (continued)
- 9. © 2018 Change Healthcare LLC and/or its subsidiaries or affiliates. All Rights Reserved. Diagnosis Probability C79.9: Secondary malignant neoplasm of unspecified site. 94.3% A81.2 Progressive multifocal leukoencephalopathy. 91.2% J44.9 Chronic obstructive pulmonary disease, unspecified. 92.3% ICD-10 Modeling
- 10. © 2019 Change Healthcare LLC and/or one of its subsidiaries. All Rights Reserved. 10 Change Healthcare AI project examples EMS Coding Radiology Coding Risk & HEDIS Coding Revenue Optimization Audit Payment Integrity Audit Full AI coding AI as a component Dual Eligibility HEDIS is a registered trademark of the National Committee for Quality Assurance (NCQA)
- 11. © 2018 Change Healthcare LLC and/or its subsidiaries or affiliates. All Rights Reserved. ● Finding codes in large charts – up to 400 pages each ● More precise labeling ● De-noising the records ● Recreating a sequence of events ● Building a language model ● Reconciling mismatch between structured and unstructured Achieving High Model Accuracy for Medical Text
- 12. © 2019 Change Healthcare LLC and/or one of its subsidiaries. All Rights Reserved. 12 AI in Healthcare Challenges Embrace Variety Respect Data Understand Workflow
- 13. © 2019 Change Healthcare LLC and/or one of its subsidiaries. All Rights Reserved. 13 “In order to understand trust in the relationship between humans and automation, we have to explore trust in two dimensions: trust in the technology and trust in the innovating firm.” Source: https://hbr.org/2017/04/to-get-consumers-to-trust-ai-show-them-its-benefits
- 14. © 2019 Change Healthcare LLC and/or one of its subsidiaries. All Rights Reserved.