Revenue-sensitive evaluation of AI-assisted ICD-10-CM coding and human-AI collaboration under dual DRG payment...

Area of Science:

• Medical Informatics
• Health Economics
• Artificial Intelligence in Healthcare

Background:

• Automated ICD-10-CM coding is vital for hospital reimbursement via Diagnosis-Related Group (DRG) systems.
• Current evaluation metrics do not differentiate error impact on revenue, potentially misrepresenting model performance.

Purpose of the Study:

• To develop and evaluate novel revenue-sensitive metrics for automated ICD-10-CM coding.
• To compare the performance of various coding models under different conditions and DRG systems.
• To assess the effectiveness of human-AI collaboration strategies guided by revenue.

Main Methods:

• Evaluated 11 models on MIMIC-IV data across diverse conditions, including full code space and zero-shot LLM samples.
• Proposed and utilized Revenue Sensitivity Index (RSI) and Coding Reimbursement Score (CRS) for evaluation.
• Simulated five human-AI review strategies and compared performance across US MS-DRG and Tw-DRG systems.

Main Results:

• PLM-ICD model achieved the highest micro-averaged F1 score (0.5934); open-source zero-shot LLMs underperformed.
• A significant 26.5% CRS gap was observed between the best and worst fine-tuned models.
• Revenue-targeted prioritization at 20% review rate reduced CRS by 43.2%, outperforming random sampling (20.0%).

Conclusions:

• Revenue-aware evaluation metrics effectively capture financially significant performance differences missed by standard metrics.
• Revenue-guided human-AI collaboration presents a promising framework for deployment, improving coding accuracy and financial outcomes.
• Model rankings remained consistent across US MS-DRG and Tw-DRG systems, indicating metric stability.