DefinePK

Abstract

Background: Artificial Intelligence (AI) is transforming respiratory care by enhancing diagnostic accuracy and streamlining workflows. The efficacy of free AI tools for spirometry interpretation, particularly in the Indian population, remains largely unassessed. This study aimed to evaluate the diagnostic accuracy of ChatGPT for spirometry interpretation compared with that of qualified respiratory physicians in a south Indian tertiary care setting.
Methods: This cross-sectional study included 100 anonymised spirometry reports that met the ATS/ERS criteria. These reports were interpreted by respiratory physicians (gold standard), and the same reports were uploaded to ChatGPT. Interpretations of the spirometry were based on ATS/ERS guidelines using Z-scores and flow-volume loops. Statistical analyses included specificity, sensitivity using proportion agreement, kappa statistics, and ROC curve analyses.
Results: The 100 spirometry reports simultaneously analysed by AI & respiratory physicians overall normal vs. abnormal classification accuracy  was 99%. For Z-score interpretation, the AI reported normal (25% vs. 26%), restriction (22% vs. 19%) obstruction (39% vs. 28%), and mixed (17% vs. 24%) compared to respiratory physician interpretation. In classifying flow-volume loop patterns, AI showed normal (30% vs 26%), restriction (18% vs. 26%) obstruction (50% vs. 38%), and mixed (2% vs. 10%) compared with respiratory physicians. In the final interpretation combining z-score & flow volume loop, AI interpretation was - 25% normal, 19% restriction, 39% obstruction, & 17% mixed, compared to the respiratory physician interpretation 26% normal, 21% restriction, 29% obstruction, & 24% mixed. AI achieved 99% agreement for normal, 98% for restriction, 90% for obstruction 93% for mixed.
Conclusion: ChatGPT is a promising tool for spirometry interpretation, but other similar AI platforms with larger samples need to be assessed before formal recommendations can be made.