| dc.description.abstract | In recent years, new technologies such as inertial measurement units (IMU) have been
introduced in applied research in sport sciences, as well as training and competition.
The use of IMU has been presented as a promising area in the sport sciences for: (1)
biomechanical analyses; (2) classification of actions and prediction of variables by
machine learning methods; and (3) quantification of training and competition physical
demand, therefore, three experiments were conducted addressing these purposes.
Experiment I – We analyzed the validity (agreement between measurement methods)
and test-retest reliability (consistency of measurement between days) of the use of
IMU when compared to reference devices, for two different tasks: (A) Standardized
vertical jumps – IMU against a force platform and a linear encoder to estimate flight
time, jump height and impulse; and (B) 30-meter sprints – IMU against a rotary encoder
(Race Encoder - RE) to estimate mean and maximum running velocity. A total of 20
individuals were recruited to perform standardized jumping tasks (squat jumps – SJ
and countermovement jumps – CMJ: 17.35 ± 2.28 years; 60.99 ± 8.87 kg; 169.30 ±
9.30 cm) and 19 individuals for sprints (17.42 ± 2.32 years; 61.45 ± 8.86 kg; 169.74 ±
9.35 cm), in two separate sessions. The tested low-cost IMU presented valid estimates
for flight time, jump height and impulse when compared to the force platform and linear
encoder, with high test-retest reliability between days for the SJ and CMJ. Additionally,
despite a significant underestimation the mean velocity of sprints, the maximum speed
estimated by UMI showed agreement when compared with the rotary encoder (Race
Encoder®). Therefore, valid estimates with acceptable repeatability were provided by
the UMI for the mean and maximum velocity of sprints. Experiment II – The
classification performance of locomotor actions common to the team sports (simulated
circuit using UMI) was evaluated using machine learning methods. For the use of
characteristics obtained from accelerometers (alone), the trained and tested classifiers
(Decision Trees, k - Nearest Neighbors, Support Vector Machines, Ensemble, Neural
Networks) showed lower detection sensitivity for moderate intensity running actions
and change of direction (COD). However, the combination of sensors (accelerometers
and gyroscopes) and aggregation of different types of extracted features (descriptive
statistics, custom statistics, and time series analysis measures) showed better overall
accuracy, sensitivity, specificity, and classification accuracy for the k-NN algorithm
“fine” (>90%), compared to the isolated use of accelerometers, in the actions of
remaining static, walking, light intensity running, moderate intensity running, sprints
with deceleration, CMJ and COD. Experiment III – To identify how the variables
obtained by UMI during handball games are grouped into factors and can be
summarized to assess the magnitude of the physical demands from the new
constructed variables. A total of 14 amateur athletes were monitored using UMI in three
consecutive handball games. The use of dimensionality reduction procedures from the
original variables quantified using UMI, showed 2 to 3 components and retained
factors, with explained variance greater than 85%. The components and factors with
the highest explained variance (> 65%) generally represent useful indicators for
monitoring and understanding the magnitude of the physical demand in handball. In
summary, the use of low-cost IMU for the analysis of standardized vertical jumps (flight
time, jump height and impulse) and sprints (maximum and mean velocity) showed
validity and repeatability when compared to the adopted reference measurements. The
trained and tested classification algorithms showed lower event identification sensitivity
for moderate running and MMD (accelerometer only). However, the combination of sensors and aggregation of extracted features showed better classification accuracy
for the k-NN algorithm (>90%) for all studied actions in both training and testing stages.
Finally, the use of UMI associated with dimensionality reduction methods showed that
the components and factors of greater explained variance allow the assessment of the
magnitude of physical demands in handball. The components and factors with the
lowest explained variance lead to an indicator for the potential occurrence of
microlesions. In addition, a set of 2 to 3 linear combinations and factors were retained.
Additionally, the weighting of the eigenvalues of the combinations retained to provide
an overview of the physical demand in handball. | |
