An autonomous bronchoscopy robot controlled by artificial intelligence (BronchoBot) outperforms experienced bronchoscopists in a simulated setting

Abstract

Background Performance in flexible bronchoscopy varies even among experienced bronchoscopists. This study aimed to compare an autonomous bronchoscopy robot (BronchoBot) to human bronchoscopists in a standardised simulation-based setting. Methods We compared 30 bronchoscopies performed by humans (10 novices, 10 intermediates and 10 experienced) to 10 bronchoscopies performed by BronchoBot on a realistic phantom. The primary outcomes were Anatomy Rating (0–12 points) and Dexterity Rating (0–6 points) assessed by two blinded expert raters. Secondary outcomes were Diagnostic Completeness (DC; 0–18 segments), Structured Progress (SP; 0–18 progressions), Procedure Time (PT; seconds) and Mean Intersegmental Time (MIT; seconds per segment), assessed by artificial intelligence (AI). Tertiary outcomes included five targeted navigations by BronchoBot to RB4→RB5→RB6 and LB8→LB9→LB10, following the bronchoscopy skills and task assessment tool (BSTAT), rated by AI. Results BronchoBot significantly outperformed all three human groups for all outcome measures. Anatomy Rating (novice, intermediate, experienced versus BronchoBot): 4.8, 10.3, 10.3 versus 11.8 points; p<0.001. Dexterity Rating: 2.0, 3.0, 4.5 versus 5.3 points; p<0.001. DC: 9, 15, 16 versus 18 segments; p<0.001, p=0.001, p=0.003. SP: 2, 7, 8 versus 18 progressions; p<0.001. PT: 281, 251, 190 versus 124 s; p<0.001. MIT: 32, 16, 13 versus 7 s per segment; p<0.001. BronchoBot successfully completed targeted navigation in all attempts for RB4→RB5→RB6: median±interquartile range 26±2 s, and LB8→LB9→LB10 30±2 s. Conclusions BronchoBot outperformed experienced human bronchoscopists using validated assessment tools based on expert human and AI ratings. BronchoBot can serve a role as benchmarking expert performance in a simulated setting, but further development and evaluation are required for use in patients.