Application Of Artificial Intelligence to Pulmonary Function Test Interpretation: A Scoping Review

Abstract

Pulmonary medicine is the field where artificial intelligence (AI) is actively used, and respiratory data can be processed objectively and in a standardized way (3). The observance of complex trends among spirometric and volumetric values that are frequently ignored by clinicians can be identified by machine learning algorithms, and AI-aided Pulmonary Function Test (PFT) interpretation is a logical development (4). Nevertheless, the interpretation of PFT is still susceptible to the high interobserver variability, and the accuracy rate of pulmonologist varies in a range of 44.6% to 65.8 percent in contrast to expert consensus (4). Although AI has the potential to perform at the same level as experts or even better, the scoping study has not been performed in a systematic manner to map the new field. This review covered studies that used adults (over 18 years) that used AI, machine learning or deep learning algorithms to interpret, classify or diagnose respiratory disease using PFT data; those using AI on non-PFT signals were not included. PubMed, Scopus, Web of Sciences, and ProQuest were searched electronically since its inception and only English-language sources were included. The review was based on the JBI scoping review methodology and PRISMA-ScR. A standardized form was used to chart the data and NVivo version 15 was used to synthesize them thematically. Out of 312 records examined, five studies were included. AI identified 40 to 100 percent, versus 44.6 percent to 74.4 percent pulmonologists (4). Explainable AI enhanced the diagnostic accuracy of clinicians by 5-10 percent, and human-AI teams had higher diagnostic accuracy compared to individual human or AI (5). Nonetheless, AI was less sensitive to restrictive patterns and rare conditions (4), no researchers conducted external validation of heterogeneous populations (4, 5), and automation bias was found. AI demonstrates a high potential in standardizing the interpretation of PFT, though without external validation, low levels of algorithmic transparency, automation bias, and absence of prospective outcome trials are still major limitations to clinical translation (3, 4, 5).