AI-Driven Radiologist-Based Reporting Optimization and Automated Medical Coding at the University of Miami Health System
Executive Summary
UHealth – University of Miami Health System sought to improve upon manual ICD-10 code verification and limited NLP accuracy in radiology reports, and to minimize delayed billing and claim rejections. Partnering with Quantiphi, UHealth and the Miller School of Medicine implemented a Generative AI–powered ICD-10 coding solution on Amazon Bedrock, combining automated code prediction with radiologist validation. This streamlined workflow is intended to mitigate radiologist burn-out and confirm their high report quality via clinical-grade coding accuracy, which will augment UHealth providing high quality care, as well as provide UHealth with the ability to prove that high value care has been delivered. This helps position UHealth as a leader in AI-driven healthcare innovation.
About the Client
The University of Miami Leonard M. Miller School of Medicine is Florida’s first medical school and with the University of Miami Health System, is a nationally recognized leader in education, research, and clinical care. With over 1,700 faculty, 48 institutes, and robust NIH funding, it’s known for breakthroughs in genomics, cancer care, and cellular therapeutics.
Problem Statement
As part of its innovation journey, UHealth identified opportunities to improve in its coding workflows—particularly in the manual verification of ICD-10 codes in radiology. Unverified codes, delayed coding, and a lack of clearly documented medical necessity reasoning can lead to billing delays, denials, or rejections, a costly industry-wide challenge. Optimally, the field of radiology could help reduce such challenges by a first of its kind approach for self-attested codes by radiologists in a quick and timely fashion, to both ensure legitimate reimbursement and that patients are followed up via their unique “No Findings Left Behind”™ provenance system-wide initiative.
To address this, UHealth radiologists sought to streamline the workflow using Generative AI to predict ICD-10 codes based on transcribed radiology reports. However, ensuring clinical-grade accuracy remained a concern due to the potential for hallucinations in generative AI outputs. The solution required a hybrid approach that combined generative AI with existing rule-based systems at UHealth and a prior innovative NLP engine (Radnosis, VEEV, Inc. Coral Gables, FL) to enhance billing accuracy, operational efficiency, and trust in AI-driven predictions.
Challenges
- Limited NLP Accuracy: Existing NLP tools lack the precision needed to interpret complex, context-rich clinical reports, leading to inconsistent ICD-10 code mapping.
- Manual Verification Bottlenecks: A reliance on manual review in high frequency could lead to slow diagnostic turnaround times and increased administrative workload.
- Lack of Transparency: Legacy systems can provide limited visibility into how codes are derived, complicating clinical validation and regulatory audits.
- Need for Understandable AI: A clear, defensible rationale for each ICD-10 prediction is essential to maintain clinician trust and meet compliance requirements.
Solution
Quantiphi partnered with the University of Miami Health System to implement a Generative AI-powered ICD-10 coding solution using Amazon Bedrock to automate code prediction from radiology reports. The solution introduced a human-in-the-loop validation framework, where AI-generated ICD-10 code suggestions were reviewed and confirmed by radiologists, ensuring clinical-grade accuracy and compliance. A custom user interface enables radiologists to efficiently validate AI predictions, significantly reducing manual effort and turnaround time.
The first phase focused on a strategic assessment to identify high-impact use cases and define an implementation roadmap, culminating in a proof of concept (PoC) to demonstrate the accuracy and reliability of the LLM-driven coding and attestation system.
By combining the strengths of large language models (LLM) and expert oversight, the solution:
- Automatically generates ICD-10 codes from clinical narratives
- Streamlines radiology workflows, reducing administrative burden
- Deploys a first of its kind code attestation by radiologists, ensuring accurate, legitimate billing for timely claims
- Expedites data outputs such as the radiologists’ codes that can be used to track/schedule ‘at-risk’ patients for follow up on actionable incidental findings (AIF’s), via UM’s Provenance network (patent-pending).
- An innovative system to alert physicians and payors of pertinent negative findings (PNF’s), to expedite care, potentially affecting patient-centric factors such as Length of Stay (LOS), etc.
Built for scalability, this GenAI framework lays a robust foundation for expanding AI capabilities across additional anatomical domains and radiology workflows. It also fosters continuous discovery of new AI use cases—positioning Uhealth Radiology at the forefront of AI-driven healthcare innovation in delivering high value care.
Technologies Used
Amazon SageMaker
Amazon Bedrock
Amazon S3
AWS DynamoDB
Amazon OpenSearch Service
AWS Lambda
AWS CloudFormation
Results and Business Impact Created
- 92% coding accuracy on real-world data, ensuring clinically reliable results.
- 70% reduction in manual effort, allowing radiologists and coders to focus on complex, high-value cases.
- Up to 8x faster processing speeds, cutting turnaround times from days to hours.
- Supports projected revenue increase, driven by the identification and correction of missing or misattributed billable conditions.
- In more than 50% of patients, it Improves the documentation of, and legitimate reimbursement for PNF’s (Pertinent Negative Findings). This expedites patient care and discharge by lowering LOS (length of stay).
- In 15-25% of patients (projected), it accelerates the follow up of AIF’s (Actionable Findings) via UM’s “No Findings Left Behind”™ provenance initiative for further, necessary imaging.
- Projected processing of over 1 million advanced imaging reports annually—including CT, MRI, and PET.
As a key Partner Innovation Alliance member of the AWS Generative AI Innovation Center (GenAIIC), Quantiphi collaborated closely with the GenAIIC team to develop this solution, benefiting from expert guidance by their Science Advisory experts. By integrating the power of large language models (LLMs) with deep scientific oversight from the GenAIIC Science Advisory team, the application will help enable UHealth-Miller School’s Department of Radiology to elevate operational excellence and accelerate its journey toward an AI-driven healthcare ecosystem.