Agent-guided AI-powered interpretation and reporting of nerve conduction studies and EMG (INSPIRE)

Objective: We aimed to create a tool for electrophysiologist enhancing and standardizing interpretation of neuromuscular electrodiagnostic tests (EDX) using state of the art generative AI technology. Methods: We developed three model frameworks for interpreting and reporting EDX: (1) Base-LLM (large language model), employing one-shot inference; (2) INSPIRE (Agent-Guided AI-Powered Interpretation and Reporting of Nerve Conduction Studies and EMG), a multi-agent AI framework; and (3) INSPIRE-Lite, a cost-efficient version of INSPIRE. INSPIRE uses three agents integrating tools to read reference tables and long-context clinical neuromuscular textbook. Performance was evaluated using the AI-Generated EMG Report Score (AIGERS), a scoring system we developed. Results: INSPIRE achieved an accuracy of 92.2 % for detecting normal versus abnormal tests, significantly outperforming the Base-LLM model, which achieved 62.6 % (p < 0.001). INSPIRE demonstrated significantly higher AIGERS scores overall and across the domains of finding, clinical diagnosis, and semantic concordance (p < 0.001). INSPIRE-Lite scored lower than INSPIRE in finding and clinical diagnosis (p = 0.001 and p = 0.004). Conclusion: Our model integrates variables like patient medical history, current complaints, and EDX findings to manage and interpret EMG. Demonstrating superior performance while addressing hallucinations, data overload, and aiding prioritization and standardization. Significance: This model enables comprehensive analysis by integrating diverse clinical variables, enhancing diagnostic accuracy and efficiency of EDX reports.