Table 2.

Comparison of AI-assisted medical coding studies.

Aspect	Study of Soroush et al. (2024)	Our study
Focus	Broad medical coding (ICD-9-CM, ICD-10-CM, and CPT)	Nephrology-specific ICD-10 coding
AI models	GPT-3.5, GPT-4, Gemini Pro, and Llama2-70b Chat	ChatGPT 3.5 and 4.0
Input format	Official code descriptions	Clinical scenarios mimicking pre-visit testing
Task	Generate exact matching codes	Identify single most appropriate ICD-10 code
Prompt design	Standardized for code generation	Simple, clinically relevant
Top performance	GPT-4: 45.9% (ICD-9-CM), 33.9% (ICD-10-CM), and 49.8% (CPT)	ChatGPT 4.0: 99% (ICD-10 for nephrology)
Performance range	Below 3% to below 50%	87–99%
Code types	Multiple (ICD-9-CM, ICD-10-CM, and CPT)	Single (ICD-10)
Specialty focus	General medical	Nephrology-specific
Main finding	Base LLMs inadequate for medical coding	AI shows high potential for specialty-specific coding.
Accuracy factors	Code frequency, length, description conciseness	Specialty focus, clinical context, latest AI versions.
Conclusion	Need for further research on complex ICD structures	AI can reduce administrative burden in specialty coding through Nephrology case scenarios for pre-visit testing.