Exercise Classification in Resistance Training: A Systematic Review of Technological Approaches

Modern sensor technology allows for objective tracking of resistance training exercises, yet the accuracy with which these technological approaches can classify which exercise is being completed is mixed. With commercially available technology commonly claiming the ability to characterise resistance training variables (e.g. exercise type and volume), synthesis of the current evidence base is warranted.

The aims of this systematic review were to (1) summarise the methodologies which have been used to achieve exercise prediction in resistance training and (2) compare the predictive performance of technologies and predictive models.

A systematic search of four databases was performed. Included studies were: development and/or validation studies; concerned with the measurement of kinetics and/or kinematics of resistance training exercises; and used statistical prediction modelling for exercise classification. The Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies (CHARMS) tool was used for data extraction, and the Prediction Model Risk of Bias Assessment Tool (PROBAST) tool was used to assess risk of bias and applicability. A total of 44 studies were included (2 validation; 42 development and validation studies).

Various technologies have been evaluated, namely: inertial measurement units, accelerometers, electromyography, electrocardiography, two-dimensional (2D) cameras, force-sensitive resistors, stretch sensors, capacitive proximity sensors, cellular signal receivers, active sonar systems, passive radio frequency identification tags, rotary encoders, and load cells. Inertial measurement units appear to be the most accurate technology available and, when worn on the wrist of the athlete, offer excellent accuracy, even for lower body exercises. Other measurement technology worn by the athlete, such as electromyography and smart materials, also offer very good accuracy. Externally placed devices, whilst offering excellent accuracy, have practical limitations that may compromise their feasibility. Of note, the exercises included the classification problem, and specifically, how similar the exercises were had a significant impact on accuracy.

Standardising the classification problem is strongly recommended as it will likely facilitate a clearer understanding of the best approach and inform consumers and future research into this area. Furthermore, ensuring technologies are robust to the prediction of a large range of exercises with similar movement patterns remains a priority and potential barrier to feasibility. Overall, accurate exercise classification is possible with sensor-based technology, although end-user availability of such technology is limited. It is strongly advised that users be cautious of consumer-level technology, because few are scientifically validated.