Artificial intelligence (AI) software for spirometry quality control: external validation in primary care

Le résumé fourni par la source

Introduction: Suboptimal quality of spirometry in primary care may contribute to misdiagnosis. Previous retrospective analyses of clinical trial data have shown that AI software has comparable accuracy as experts when assessing spirometry quality (Topole et al ERJ Open Res 2023). Aim: To validate the performance of AI software in assessing technical quality of primary care spirometry against expert respiratory physiologists according to 2019 international standards. Methods: Two hundred consecutive spirometry sessions, conducted in a primary care diagnostic hub, were independently assessed by two blinded expert respiratory physiologists, and by AI software (ArtiQ, Belgium). For each curve, FEV1 and FVC for the whole session were graded according to ATS/ERS 2019 technical standards (A, B, C, D, E, F, U). We classified grades A or B (at least two acceptable traces within 0.150L) as “good” quality, and other grades as “suboptimal” quality with the reference standard being expert consensus. Results: Experts had good and moderate interrater agreement on assessing the quality of FEV1 (Κ = 0.62) and FVC (Κ = 0.51) with consensus for both FEV1 and FVC in 83.5% sessions. AI software agreed with expert consensus for 88.0% and 83.8% of FEV1 and FVC sessions respectively. AI had a sensitivity, specificity and positive predictive value of 89.9% (95% CI 84.3-94.0), 68.8% (95% CI 41.3-89.0) and 96.8% (95% CI 93.6-98.4) for FEV1, and 88.3% (95% CI 82.2-92.9), 47.4% (95% CI 24.5-71.1) and 93.2% (95% CI 89.8-95.4) for FVC. Conclusion: AI software had comparable performance as individual experts. There is potential for AI software to automate quality assessment of primary care spirometry.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.