The results from the calculation of extracted features for
the different severities of resting and postural tremor are
listed in Fig. 5. It is shown in the figure that the value of
every feature increases as tremor severity increases. The first
two and second two columns depict resting and postural
tremor results, respectively. The position of the sensor
(index finger or wrist) alternates with each column starting
with the finger position. Rows one to three are gyroscope
feature results: 1. Natural logarithm of the root mean square
of angular velocity; Log(RMS(Ang. Vel.)) 2. Natural
logarithm of gyroscope peak power; Log(Gyro Power) 3.
The standard deviation of angular velocity; Stdev(Ang.
Vel.). Rows four to six are accelerometer feature results: 4.
Natural logarithm of the root mean square of linear
acceleration; Log(RMS(L. Accel.)) 5. Natural logarithm of
accelerometer peak power; Log(Accel Power) 6. The
standard deviation of linear acceleration; Stdev(L. Accel.).
The classification accuracy of the resting and postural
tremor severity quantification models for different
combinations of features are shown in Table I. The different
combinations of features were selected based on the
placement of sensor on the body (index finger and wrist) as
well as sensor type (gyroscope and accelerometer). For
example, when the features calculated from both sensors
fixed on the index finger were used to train the classifier, it
achieved a classification accuracy of 88.9% (column 3, row
4) for resting tremor and 82.1% (column 4, row 4) for
postural tremor. However, when the classifier was trained
with all the features calculated from both sensors on both the
index finger and wrist, the accuracy was 83.3% (column 3,
row 10) for resting tremor and 79.5% (column 4, row 10) for
postural tremor. In total, the classifier performance for all
the sensors and locations were computed separately and
together (Table I).
The results from the calculation of extracted features forthe different severities of resting and postural tremor arelisted in Fig. 5. It is shown in the figure that the value ofevery feature increases as tremor severity increases. The firsttwo and second two columns depict resting and posturaltremor results, respectively. The position of the sensor(index finger or wrist) alternates with each column startingwith the finger position. Rows one to three are gyroscopefeature results: 1. Natural logarithm of the root mean squareof angular velocity; Log(RMS(Ang. Vel.)) 2. Naturallogarithm of gyroscope peak power; Log(Gyro Power) 3.The standard deviation of angular velocity; Stdev(Ang.Vel.). Rows four to six are accelerometer feature results: 4.Natural logarithm of the root mean square of linearacceleration; Log(RMS(L. Accel.)) 5. Natural logarithm ofaccelerometer peak power; Log(Accel Power) 6. Thestandard deviation of linear acceleration; Stdev(L. Accel.). The classification accuracy of the resting and posturaltremor severity quantification models for differentcombinations of features are shown in Table I. The differentcombinations of features were selected based on theplacement of sensor on the body (index finger and wrist) aswell as sensor type (gyroscope and accelerometer). Forexample, when the features calculated from both sensorsfixed on the index finger were used to train the classifier, itachieved a classification accuracy of 88.9% (column 3, row4) for resting tremor and 82.1% (column 4, row 4) forpostural tremor. However, when the classifier was trainedwith all the features calculated from both sensors on both theindex finger and wrist, the accuracy was 83.3% (column 3,row 10) for resting tremor and 79.5% (column 4, row 10) forpostural tremor. In total, the classifier performance for allthe sensors and locations were computed separately andtogether (Table I).
การแปล กรุณารอสักครู่..
