AI-Assisted Interpretation of Pulmonary Function Tests: Enhancing Diagnostic Precision in Respiratory Medicine Authors: Ajay Singh, Manoj Sahu, Surendra K. Padarya, Talha Saad, Satyendra Mishra, Mayank P. Singh Published: 2025 ABSTRACT Background:Interpreting pulmonary function tests (PFTs) accurately is essential for diagnosing and monitoring respiratory disorders. However, manual interpretation is subject to interobserver variability. Artificial intelligence (AI) offers a promising tool to enhance diagnostic consistency and accuracy.Objectives:To evaluate the diagnostic accuracy of an AI-based interpretation system for PFTs in comparison with expert pulmonologist consensus and to assess its applicability in clinical settings.Methods:A cross-sectional diagnostic accuracy study was conducted among 200 adult patients undergoing spirometry and full PFT. AI software analyzed raw PFT data to classify patterns as normal, obstructive, restrictive, or mixed. Three pulmonologists independently reviewed the same data, and their consensus was considered the reference standard. Diagnostic performance metrics, including sensitivity, specificity, accuracy, and Cohen’s kappa, were calculated.Results:The AI system correctly identified 38/40 normal, 76/80 obstructive, 44/50 restrictive, and 27/30 mixed patterns, with an overall diagnostic accuracy of 92%. Sensitivity and specificity were 91.5% and 93.2%, respectively. Cohen’s kappa was 0.86, indicating strong agreement with expert interpretations.Conclusion:AI-assisted interpretation of PFTs demonstrates high diagnostic concordance with pulmonologist consensus and holds potential for standardized, efficient respiratory diagnostics. INTRODUCTION Pulmonary function tests (PFTs) are vital tools in the evaluation and monitoring of respiratory diseases, offering objective measurements of lung volumes, flow rates, and gas exchange. They are essential for diagnosing conditions, such as asthma, chronic obstructive pulmonary disease (COPD), restrictive lung diseases, and preoperative pulmonary risk stratification. However, interpreting PFT results requires integration of complex spirometric patterns with clinical context, which can be challenging even for experienced clinicians. Interobserver variability and misclassification of ventilatory defects continue to be a concern in respiratory diagnostics. Recent advancements in artificial intelligence (AI) and machine learning (ML) have opened new avenues for enhancing diagnostic accuracy across multiple domains in medicine, including respiratory care. AI-based algorithms, trained on large datasets, can identify subtle patterns in PFT data that may be overlooked by traditional interpretation methods. These tools offer the potential to standardize interpretation, reduce diagnostic errors, and provide decision support in real-time clinical settings.[12] Studies have shown that AI systems can classify PFT results into obstructive, restrictive, or mixed ventilatory defects with accuracy comparable to that of pulmonologists, and in some cases, even outperform them in specific diagnostic tasks.[3] Moreover, AI-assisted systems can integrate additional clinical variables, such as age, sex, smoking history, and symptoms, to generate context-sensitive diagnostic suggestions, thereby personalizing respiratory assessment. Despite these promising developments, the implementation of AI in routine pulmonary practice remains limited due to concerns about transparency, validation across diverse populations, and integration with existing clinical workflows. This study aims to evaluate the diagnostic performance and clinical utility of an AI-assisted interpretation system for PFTs in a tertiary care setting and compare its accuracy with that of human experts. METHODOLOGY This was a cross-sectional, diagnostic accuracy study conducted in the Department of Pulmonary Medicine at a tertiary care teaching hospital over a six-month period. The study aimed to compare the interpretative accuracy of an artificial intelligence (AI)-assisted pulmonary function test (PFT) analysis system with that of board-certified pulmonologists. Ethical considerations Institutional Ethics Committee approval was obtained prior to study commencement. Written informed consent was secured from all participants. All data were handled confidentially in compliance with institutional and national data protection policies. Study population A total of 200 adult patients (age ≥18 years) who underwent spirometry and full pulmonary function testing as part of routine clinical care were included. Patients with poor quality spirometry (per ATS/ERS standards), incomplete demographic data, or known neuromuscular disorders affecting respiration were excluded. Data collection Demographic details, clinical history, and PFT parameters, including FEV1, FVC, FEV1/FVC ratio, PEF, TLC, and DLCO, were recorded. Each patient’s data were anonymized and uploaded into the AI system (trained using supervised learning techniques on a dataset of >10,000 labeled PFTs). The AI model was programmed to classify each case as normal, obstructive, restrictive, or mixed ventilatory defect. Comparator Three experienced pulmonologists, blinded to AI results, independently interpreted the same PFT reports using standard interpretation guidelines (ATS/ERS 2019). Their consensus diagnosis served as the reference standard. Outcomes and analysis The primary outcome was diagnostic concordance between AI interpretation and expert consensus. Sensitivity, specificity, accuracy, and Cohen’s kappa coefficient were calculated to assess agreement. Subgroup analysis was performed based on diagnosis type (obstructive vs. restrictive), patient age, and spirometry quality grade. RESULTS Out of 200 participants, expert consensus classified 40 cases as normal, 80 as obstructive, 50 as restrictive, and 30 as mixed ventilatory defects. The AI system correctly identified 38 normal, 76 obstructive, 44 restrictive, and 27 mixed patterns, achieving diagnostic accuracy rates of 95%, 95%, 88%, and 90%, respectively, as seen in Table 1. Table 1Diagnostic classification accuracy by AI vs. expert consensusDiagnosis typeExpert diagnoses (n)AI correctly identified (n)Accuracy (%)Normal403895%Obstructive807695%Restrictive504488%Mixed302790%AI=Artificial intelligence Diagnostic classification accuracy by AI vs. expert consensus AI=Artificial intelligence Overall, the AI system demonstrated a sensitivity of 91.5%, specificity of 93.2%, positive predictive value (PPV) of 92.1%, and negative predictive value (NPV) of 92.6%. The overall diagnostic accuracy was 92.0%, and inter-rater reliability assessed by Cohen’s kappa was 0.86, indicating strong agreement with pulmonologist interpretations, as seen in Table 2. Table 2Summary of AI diagnostic performance metricsMetricValueSensitivity91.5%Specificity93.2%Positive predictive value (PPV)92.1%Negative predictive value (NPV)92.6%Overall accuracy92.0%Cohen’s kappa0.86AI=Artificial intelligence Summary of AI diagnostic performance metrics AI=Artificial intelligence These findings support the utility of AI in standardizing pulmonary function test interpretation and reducing observer variability. DISCUSSION The results of this study demonstrate that AI-assisted interpretation of pulmonary function tests (PFTs) offers high diagnostic accuracy, closely aligning with the consensus of expert pulmonologists. With an overall accuracy of 92% and a Cohen’s kappa value of 0.86, the AI system in this study proved to be a reliable tool in identifying ventilatory defects, including obstructive, restrictive, and mixed patterns. Similar findings have been reported in earlier studies. Topalovic et al.[1] showed that AI could outperform pulmonologists in PFT interpretation, particularly in complex cases, reducing diagnostic variability and increasing standardization. Our results further confirm the ability of AI to replicate expert-level reasoning using structured data inputs, such as FEV1, FVC, and DLCO, aligning with prior work by Moor et al.,[3] which highlighted early prediction capabilities in lung disease using machine learning models. The AI system demonstrated high specificity and negative predictive value, indicating its potential utility in clinical triaging and early identification of normal patterns to reduce unnecessary referrals.[4] In settings with a shortage of trained respiratory specialists, such technology could support primary care physicians and optimize workflow.[56] One notable observation was the slightly lower sensitivity in restrictive patterns (88%), a limitation previously highlighted in algorithm-based assessments where nuanced volume-based defects pose a challenge.[7] Continuous training on larger and more diverse datasets may address this limitation. Despite its promise, the integration of AI tools into clinical practice faces barriers, such as provider trust, algorithm transparency, and the need for rigorous validation across populations.[89] Our study adds to the growing body of evidence advocating for AI adoption in respiratory diagnostics, especially in resource-constrained environments. CONCLUSION This study demonstrates that AI-assisted interpretation of pulmonary function tests can achieve diagnostic accuracy comparable to expert pulmonologists, with strong agreement across major ventilatory patterns. The technology offers a valuable decision-support tool, particularly in resource-limited settings or high-volume clinical environments. While the current AI model performed well, especially in identifying obstructive and normal patterns, further refinement is needed to enhance sensitivity in detecting restrictive defects. Overall, AI integration holds promise for improving the efficiency, consistency, and accessibility of respiratory diagnostics. Conflicts of interest There are no conflicts of interest.