Background: Telerehabilitation can improve access to rehabilitation services for older adults, but limited clinical supervision may increase the risk of undetected compensatory movements. This study evaluated whether inertial measurement unit (IMU) based machine learning model could automatically detect compensations during rehabilitation exercises. Methods: Five exercise-specific classifiers were developed using a controlled dataset of compensatory movements performed by fifteen healthy kinesiologists. For each exercise, 450 correct and 450 compensatory repetitions were collected while participants wore both the prototype and commercial IMUs. Biomechanical and statistical features were extracted, and LASSO regression was applied for feature selection. Five different classifiers were optimized through grid search and evaluated using subject-wise leave-one-out cross-validation. The selected models were subsequently validated in eleven community-dwelling older adults aged 69.7 ± 7.1 years. Results: In controlled conditions, the prototype achieved performance comparable to the commercial IMU (accuracy and precision ≥ 95%). In older adults, accuracy ranged from 67% to 96%, whereas precision ranged from 32% to 100%, indicating exercise-dependent differences. Conclusions: A low-cost IMU combined with lightweight machine learning can detect compensatory movements with performance comparable to a research-grade sensor under controlled conditions. Larger clinical samples, different exercises and multi-level compensation labels are required to improve generalizability and support real-time telerehabilitation feedback.