ChatGPT-assisted pain history-taking system for musculoskeletal rehabilitation: Development and preliminary psychometric validation.

BACKGROUND

Chronic musculoskeletal pain requires comprehensive history-taking, yet traditional approaches often yield inconsistent documentation and limited cross-visit comparability. Evidence for evaluated artificial intelligence (AI)-assisted pain history-taking tools remains limited.

OBJECTIVE

To develop and preliminarily evaluate the Chronic Musculoskeletal Pain-E-history Acquisition System (CMP-EAS), a ChatGPT-assisted, framework-anchored structured workflow for pain history-taking.

METHODS

Thirty-six patients with chronic musculoskeletal pain completed CMP-EAS interviews in a single-center observational study; five experts assessed content validity. Content validity was assessed using item- and scale-level indices (I-CVI, S-CVI/Ave), with relevance and clarity evaluated separately. Deterministic repeatability under controlled conditions was examined in three independent retests using agreement and prevalence- and bias-adjusted kappa (PABAK) at both item and response levels. User experience was assessed using a 21-item survey (5-point Likert scale).

RESULTS

CMP-EAS comprised 13 items covering core pain-history domains. Content validity was excellent (S-CVI/Ave = 1.00 for relevance; 0.97 for clarity). Item-level repeatability was perfect (agreement, 100%; PABAK = 1.00), and response-level repeatability was near-perfect (agreement, 94-100%; PABAK = 0.89-1.00). User experience was favorable (mean ± SD 4.2 ± 0.5/5.0), with interaction comfort rated highest (4.4 ± 0.5) and trust lowest (4.0 ± 0.7).

CONCLUSION

In a Mandarin-speaking sample from a single tertiary center in southern Taiwan, CMP-EAS demonstrated preliminary content validity, structural consistency, high deterministic repeatability, and favorable user experience. A one-question-at-a-time, framework-anchored workflow may standardize initial pain-history documentation and support cross-visit comparison. These findings reflect system-level reproducibility rather than real-world clinical performance; broader validation is required before clinical generalization.