Aryan Sadghian, Marc Mitchell

Background: Canadians struggle to meet resistance training (RT) guidelines despite clear health benefits. Smartphone-based augmented reality (AR) motion capture may offer a scalable self-monitoring solution, but its pragmatic validity versus human counting under real-world conditions remains under-tested.Methods: This ongoing validation study compares a deep learning-powered AR rep-counting tool against manual counting. The planned sample is 26 adults (13 students, 13 office workers). To date, the student cohort is complete (n=13). Participants performed 10-repetition sets of squats, sit-ups, push-ups/knee push-ups, and high knees under two clothing conditions (form-fitting; loose long-sleeve top/pants). Each set was captured while the AR tool ran live and recorded for later human scoring. Two independent raters counted repetitions from video; their mean served as the criterion. We computed set-level error (AR – manual), mean absolute error (MAE), and proportions within ±1 and ±2 repetitions. Equivalence of mean counts was tested using paired two one-sided tests (TOST) with bounds of ±1 repetition (α=.05; 90% CI).Results: Across 104 exercise sets from 13 students, inter-rater agreement was high (96.2% exact agreement; ICC(2,1)=0.95, 95% CI 0.93-0.97). The AR tool showed minimal bias relative to manual counting (mean difference=0.03 repetitions; SD=1.60) and MAE=0.59 repetitions. Overall, 88.5% of sets were within ±1 repetition and 96.2% within ±2. TOST supported equivalence (90% CI −0.23 to 0.29; both one-sided tests p<.001).Conclusions: In the student cohort, the AR rep-counting tool closely matched human counting and met a ±1 repetition equivalence margin, supporting potential replacement of manual counting for RT self-monitoring.