Decomposing the Problem: Fine-Tuning the Model
Explore how to decompose a hospital billing accuracy problem through fine-tuning AI models on labeled clinical data. Understand the constraints from HIPAA, the importance of data quality, and how to integrate AI suggestions into coder workflows within the hospital's EHR system.
This lesson works through a case study from a regional hospital that wants to improve the accuracy of its medical billing process using AI. We start by reading the full case study, then work through it step by step.
The case study
This is the complete case study. Read it once before we begin decomposing it.
The case study
The organization: A large regional hospital in the Midwest with multiple outpatient clinics. A team of medical coders assigns billing codes across all facilities.
The problem: After each patient visit, a physician writes a clinical note describing what the patient presented with, what was assessed, and what was done. Medical coders read those notes and assign two types of billing codes:
ICD-10 codes (International Classification of Diseases, 10th revision) identify the patient’s diagnosis.
CPT codes (Current Procedural Terminology) identify the procedures and services the physician performed.
Insurance companies use these codes to determine how much to reimburse the hospital. If the codes are wrong, the insurer denies the claim. The coder must then review the original note, identify the error, correct the codes, and resubmit. The hospital’s first-pass acceptance rate, the percentage of claims the insurer accepts on the first submission without denial, currently sits at 81%. The industry benchmark is 95%.
The systems: Physicians write clinical notes inside Epic, the hospital’s ...