The research field of human activity recognition by means of commercially available, wearable technologies has gained an increasing focus in sports and health science for proactively monitoring and assisting users in their activities [1
]. Wireless technologies, including Inertial Measurement Units (IMUs) and Global Positioning System (GPS) trackers, have become readily accessible in ubiquitous devices such as smartphones and smartwatches to monitor physical activity and performance in sports [2
]. Thereby, the computational power of smartphones and smartwatches is ever increasing, with enhanced user interfaces that enable analysis of the wireless data in real time [5
The application of wearable technologies to repetitive aerobic activities is well researched and successfully introduced to the market [1
]; yet, their application to resistance training still remains limited [2
]. In comparison to aerobic activities, such as outdoor running or cycling, performance monitoring of stationary strength training workouts requires careful consideration of sensor positioning and more advanced numerical analysis of the available data [3
]. In addition, the execution diversity between exercises and individual athletes further complicates the analysis [4
In an early effort to use smartphones for strength training monitoring, a dynamic time warping-based algorithm was introduced to identify exercise and count repetitions based on the available acceleration data [1
]. The proposed numerical algorithm was tested both indoors with weight machines and for outdoor scenarios using free weights and resistance band exercises with promising results, i.e., below 1% classification error rate while remaining computationally inexpensive. In similar research, a prototypical machine learning algorithm was introduced for exercise recognition of three different strength exercises with dumbbells using a wrist-worn smartwatch with a demonstrated mean recognition rate of 97.7% in 20 adults [1
]. More recent efforts led to the FitCoach, a virtual fitness coach to assess dynamic postures during workouts using data from wearables and smartphones, which was tested in 12 participants and 9 different strength exercises with an average exercise detection rate of 95% [5
]. FitCoach was developed to combine exercise recognition and interpretation of wireless data into an easy-to-understand exercise review score for performance evaluation and recommendation to avoid injury [6
]; however, no reference was made with regards to the One Repetition Maximum (1RM) as the key indicator of strength training performance.
The 1RM as ‘the maximum load that can be lifted through a full range of motion’ is known as the most valid indicator of an individual’s dynamic strength [7
], and thus, the quantification of an individual’s 1RM is fundamental in the design of safe and effective resistance training programs [8
]. The direct assessment of 1RM is time-consuming and depends on the athlete’s experience, motivation and fatigue, with risk of musculoskeletal injury due to maximum loading [9
]. In contrast, indirect methods have been introduced to predict the 1RM based on well-established linear regression techniques, including the repetition to failure method [10
] as well as the relationship between load and lifting velocity (L-V relationship) [7
]. In order to derive the L-V relationship for individual athletes and exercises, commercially available Linear Position Transducers (LPT) are generally used [16
]. Yet, the application of LPT devices to free weight and sport-specific strength exercises is compromised. In particular, LPT devices are limited in picking up fluctuations in lifting velocities due to horizontal or asymmetrical displacements depending on the positioning and manufacturer of the device [7
Recent advances in smartwatch-based technologies hold great potential to help improve 1RM predictions for strength exercises without Smith machines in the strength training-specific setting. Towards this goal, Lorenzetti and Huber [18
] introduced an iOS workout analysis application for the Apple Watch called StrengthControl to determine exercise recognition and repetition count, and, piloting towards the prediction of 1RM, muscle loading and fatigue. The StrengthControl app was tested in one subject for five resistance training exercises (barbell biceps curl, barbell bench press, barbell back squat, dumbbell lateral raise, and dumbbell biceps curl with twist), with a reported mean error in exercise recognition of 3.5% and 0.92% in repetition counting, respectively [18
]. However, no study has yet to report on the reliability and accuracy of smartwatch-based measurements in predicting 1RM outside of the research setting. The goal of this study was to assess the validity, reliability and accuracy of the iOS StrengthControl app in exercise recognition, repetition count, and 1RM prediction in recreational athletes in the strength training-specific environment.
The present results suggest that further investigations are needed to improve the accuracy of the velocity estimates from smartwatch-based readings to predict 1RM. A reduction in technical errors and time lag in data transfer may be achieved by accounting for subject-specific body height and range of motion, as well as giving clear instructions on pauses between concentric and eccentric movement phases. Future research should also consider alternative motion sensors or vision-based methods for human activity recognition to assess an individual’s 1RM in the strength training-specific setting.
Inaccuracies in exercise recognition and repetition count, as well as failed attempts to predict 1RM using the StrengthControl app, can largely be explained by inter- and intra-subject differences in exercise execution within and between sets, as well as technical difficulties with the smartwatch not being able to capture and process the data correctly. In order to execute the strength exercises with maximal concentric velocity, the participants performed rapid movements without any instructions regarding the pauses between the concentric and eccentric phase of each repetition. It was previously suggested that imposing a pause between eccentric and concentric movements would increase the reliability of acceleration measurements using a smartwatch [30
]. Thus, it is possible that clear instructions to the pauses may have helped to lower the coefficient of variation in the smartwatch data readings, thereby increasing the accuracy in exercise recognition, repetition counting and successful attempts to predict 1RM.
Technical difficulties and disturbances arose in the wireless transfer of data from the smartwatch to the smartphone. Unfortunately, the smartwatch ended up either stuck in a loop, or not all data was transmitted due to a lag in transmission. The lag was likely caused by the slow processer that is embedded in the first generation of the Apple Watch Sport. Here, a newer model of the Apple Watch may have helped to eliminate problems with wireless data transfer. However, similar research also reported that smartphone-based accelerometers presented with a considerable loss of data that was not correctly detected by the sensor during bench press exercises with the Smith machine [17
]. In contrast to accelerometers that are specifically built for high-velocity measurements with sampling frequencies of 200 to 500 Hz, accelerometers embedded in the smartwatch or smartphone remain low-cost and based on low frequency sampling that is not precise enough to analyse explosive movement and repetitive movements at higher velocities.
In comparison to direct 1RM assessment, the prediction of 1RM based on the L-V relationship can be done on a regular basis without the high risk of injury associated with maximal loading. Indeed, previous findings suggest that there is no need to test overly heavy loads, as the prediction of 1RM from the L-V relationship derived at sub-maximal loads with exercise execution at maximal velocity is just as accurate [13
]. Yet, the key challenge in the prediction of 1RM based on the L-V relationship is the requirement for accurate velocity measures during exercise performance, which were not shown to be reliable enough using the proposed methodology. Here, Peláez Barrajón and San Juan [17
] also concluded that smartphone-based accelerometers are less reliable for the measurement of concentric mean velocity during bench press exercises compared to LPT devices. It was suggested that accurate measures of range of motion and body height are required to improve the accuracy in the calculation of velocity parameters using the smartwatch [17
]. Unfortunately, subject-specific differences in body height and range of motion could not be accounted for in the StrengthControl app, likely contributing to some of the inaccuracies in the present results. Furthermore, participants may have performed strength exercises with submaximal velocity even though maximal velocity is required to adequately calculate 1RM using the L-V relationship. Thus, the option to calculate 1RM using the repetition to failure method [10
] should also be considered for implementation into any workout analysis application using wearable technologies.
As an alternative to IMUs and GPS trackers for performance tracking, research in human activity recognition has been directed towards 3D pose estimation using data from the high-speed camera in smartphones in combination with advanced image analysis and deep learning techniques in computer vision [31
]. These advances in computer vision provide alternative, and possibly complementary, means to assess lifting velocity during strength exercises for predicting 1RM. Here, the so-called PowerLift
application for the iOS was recently introduced to measure barbell velocity by video-recording the lift using an iPhone [32
]. It was demonstrated that the PowerLift
application achieved accurate and reliable mean barbell velocity measures during the full squat, bench press and hip thrust exercises when compared with the results from an LPT device [32
]. Furthermore, a method was introduced that combined a single hand-held camera and a set of 13 IMUs attached to the body limbs to estimate 3D pose in the wild [34
]. While the use of 13 smartwatch-based IMUs is not feasible for widespread application, combining smartwatch-based and iPhone-based readings with advanced deep learning techniques seems promising to open new perspectives for the advancement of strength training monitoring.
Two limitations of the present study that haven’t been discussed are the heterogeneity of participants, as well as the lack of directly assessing each participant’s 1RM for comparison with the adopted 1RM equations as the gold standard. The study group was chosen to represent potential end-users of the StrengthControl app who are common in the recreational strength training-specific setting. However, a more confined study group, for example focusing on experienced power-oriented athletes of similar age and gender, would have likely led to reduced inaccuracies in results due to smaller inter- and intra-subject differences in exercise execution. Unfortunately, directly assessing 1RM in the present study group was not feasible due to the study design, time constraints and experience of the participants. Here, power athletes may be more willing and experienced with direct 1RM testing, and should be considered for further validation of 1RM predictions from wearable and smartphone-based technology in future work.