Document resource
Background The quality of spirometry interpretation is highly variable in primary care, contributing to delayed diagnosis of chronic respiratory diseases. We hypothesised that Artificial intelligence (AI) decision support software would improve spirometry interpretation performed by primary care clinicians.Methods We conducted a parallel group, randomised controlled superiority trial. Clinicians working in primary care who routinely refer for, perform or interpret spirometry assessed fifty real-world patient spirometry records cases, providing ‘preferred diagnosis’ (most likely diagnosis) through an online platform either with (Intervention) or without (Control) AI decision support software. The primary outcome was the percentage of cases where ‘preferred diagnosis’ agreed with the reference diagnosis pre-determined by expert pulmonologists with access to full patient records. Sensitivity analysis for missing data, and subgroup analysis in cases with a reference diagnosis of chronic obstructive pulmonary disease (COPD) was performed. Secondary outcomes included performance in differential diagnosis prediction, technical quality assessment, pattern interpretation and self-rated confidence in spirometry interpretation.Results Between June 2023 and April 2024, 234 participants were randomised with 133 (57%) completing (Intervention n=67, Control n=66): 73% female, 42% were general practitioners, 50% nurses. The addition of AI decision support software improved the mean ‘preferred diagnosis’ prediction performance from 49.7% (Control) to 58.7% (Intervention) in all cases (difference 9.0 [95% confidence interval 4.5 to 13.3]%, P=0.001) and in COPD cases (15.9 [9.0 to 22.7]%, p<0.001, [ figure 1]). Sensitivity analysis showed the results were robust to adjusting for job role, sex, spirometry role and registration for spirometry accreditation, and assessing the potential impact of missing data. Differential diagnosis prediction performance and technical quality assessment performance also improved with intervention, but not pattern interpretation nor clinician confidence levels.Abstract S11 Figure 1Estimates of study group effects on primary outcome: preferred diagnosis prediction performance. Data expressed as mean difference (95% confidence interval)Conclusion In primary care clinicians, AI spirometry decision support software significantly improved diagnosis prediction performance. AI decision support can help to address suboptimal interpretation of spirometry in primary care.