Dual Component Framework Approach for Automated Nerve Conduction Study Interpretation: Rule‐Based Coding and Fine‐Tuned Language Model

INTRODUCTION/AIMS: Nerve conduction study (NCS) interpretation is labor-intensive, and automation may improve clinical workflow. We developed a dual framework integrating rule-based coding and fine-tuned language models to standardize reporting and enhance diagnostic accuracy. METHODS: The framework combined rule-based coding for converting raw NCS data into structured reports with demyelination parameter evaluation, followed by GPT-4o fine-tuning with structured NCS reports as input and ground-truth diagnostic labels as output (1335 cases, 18 diagnostic categories). On 66 test cases, diagnostic conclusions were compared against six foundation large language models (rated by junior neurophysiologists) and four physicians (rated by senior neurophysiologists). Diagnostic conclusions (1 = incorrect, 5 = correct major and minor diagnoses) and clinical rationale quality (1 = flawed, 5 = correct diagnostic flow and information) were evaluated using 5-point Likert scales. Per-category diagnostic accuracy was also evaluated. RESULTS: The fine-tuned model achieved diagnostic accuracy comparable to junior neurophysiologists (4.40 [95% CI: 4.23-4.57] vs. 4.42 [95% CI: 4.29-4.54]), outperforming zero-shot GPT-4o (fine-tuned: 4.29 [95% CI: 4.14-4.43] vs. zero-shot: 3.56 [95% CI: 3.36-3.76]). Per-category evaluation yielded macro F1 0.77 and micro F1 0.8. Clinical rationale quality scored 4.34 [95% CI: 4.16-4.52] out of 5, indicating correct diagnostic flow with minor information gaps. Conclusion generation was 16-fold faster than the best physician (2.88 vs. 46.34 s per case). DISCUSSION: This dual framework demonstrates performance comparable to junior neurophysiologists on internal test data, with robust accuracy across major diagnostic categories, streamlining complex parameter calculations and generating diagnostic reports with clinical rationales. Future directions include multicenter and external validation.