Objective:

Having shown GPT-4’s potential for localizing central nervous system (CNS) lesions, we  investigated localization capabilities of GPT-4  in the peripheral nervous system (PNS).

Background:

Localization of PNS lesions is challenging for non-neurologist healthcare providers, unfamiliar with the variant mappings of nerve roots, upper and lower plexus, and branches of peripheral nerve.

Design/Methods:

Following our earlier study methods, 51 suitably-detailed open-access case reports were utilized. Unedited text of the history and physical examination (H&P) was manually extracted. Each case input was then presented to GPT-4 in three separate trials to identify consistency of response.  We employed our previous prompt engineering, combining: 1. Chain-of-Thought (CoT); 2. Text Classification (TC); 3. Few-shot; 4. Instructional prompting. Statistical analysis identified system specificity, sensitivity, precision, and F1-score.

Results:

Overall statistics showed: sensitivity 0.97, specificity 0.84, precision 0.97, F1 score 0.97. GPT-4 lesion localization F1 score was 0.98 for spinal cord lesions, 0.97 for spinal root, 0.98 for plexus, 0.91 for peripheral nerve, 0.98 for neuromuscular junction, and 0.96 for muscle. Detailed accuracy for each peripheral lesion localization (e.g., spinal root level) showed 0.60 for the peripheral nerve, 0.72 for the plexus, and 0.86 for the spinal root lesions. Error analysis was mostly due to inaccurate anatomical knowledge, lack of thoroughness, and ignoring relevant neurological findings.

Conclusions:

Combining results from the previous study of CNS lesion localization through GPT-4 and our study for PNS localization with adequate ability to localize, this could become a useful tool in the clinical setting.