**6. Machine learning classification with the smartphone and portable media device as wireless accelerometer and gyroscope platforms**

The synergy of devices, such as smartphones and portable media devices, for quantifying human movement and machine learning offers the exciting opportunity of highly objective assessment, automated computer aided diagnosis, and prognostic forecasting. LeMoyne and Mastroianni have recently emphasized the classification of movement features of the hemiplegic affected and unaffected side. Preliminary test and evaluation has demonstrated considerable classification accuracy [1, 2, 22].

These accomplishments have featured the use of Waikato Environment for Knowledge Analysis (WEKA). WEKA is equipped with a considerable array of machine learning algorithms, such as the support vector machine, multilayer perceptron neural network, J-48 decision tree, logistic regression, and K-nearest neighbors. The appropriate machine learning algorithm is selected at the discretion of the research team with consideration to the classification endeavor at hand. The feature set that comprises the data is organized into an Attribute-Relation File Format (ARFF) file. The ARFF file is derived primarily from a series of numeric attributes that appropriately quantify the classes to be distinguished [40–42].

The ability for even a skilled team of clinicians to conclusively decide upon the presence of a hemiplegic reflex pair is a subject of contention [19, 36]. Considerable classification accuracy was attained through the amalgamation of a portable media device as a wireless accelerometer platform, a potential energy impact pendulum, and a machine learning algorithm. Each reflex response sample of the affected hemiplegic reflex response and unaffected reflex response was transmitted through local wireless connectivity as email attachments for post-processing. Software automation consolidated the trial data into a cohesive feature set. The selected machine learning algorithm was the support vector machine. This research demonstrates the considerable promise of the integration of smartphones and portable media devices as wearable and wireless inertial sensor platforms in combination with machine learning algorithms for classification of perceptively distinguishable human movement features [43].

Further investigation regarding the utility of the portable media device as a wireless accelerometer platform extended to the domain of assistive devices for gait, such as a cane. An issue with using a cane involves the concern that the patient is properly using the cane per the advice of a skilled therapist. However, such instruction is only realistically feasible respective of a relatively brief clinical appointment. A portable media device functioning as a wireless accelerometer platform can monitor the usage of a cane while the subject is walking, and the data package can be conveyed by local wireless connectivity to the Internet as an email attachment. Considerable classification accuracy was attained using logistic regression as a machine learning algorithm to distinguish between correct and incorrect cane usage scenarios with the recorded accelerometer signal deriving the feature set [44].

LeMoyne and Mastroianni demonstrated the potential of integrating a portable media device as a wireless gyroscope platform for quantifying the patellar tendon reflex response and machine learning for the classification of a hemiplegic reflex pair. The tendon reflex was elicited through supramaximal stimulation from a manually operated reflex hammer. The portable media device functioning as a wireless gyroscope platform was mounted proximal to the lateral malleolus as demonstrated in **Figure 1** [45].

In order to consolidate the affected leg and unaffected leg reflex response into a feature set their respective gyroscope signals needed to be considered, such as presented in **Figure 2**. Three numeric attributes comprehensively characterize the nature of the reflex response:


These three numeric attributes represent the reflex response in terms of both kinematic and temporal features. The feature set was evaluated through WEKA using the J48 decision tree. An advantage of the WEKA J48 decision tree is the ability to visualize the decision tree, which may infer the most predominant aspects of the feature set [45].

The J48 decision tree is illustrated in **Figure3**. Note that the feature set attribute TimeMaxMinGamma is clearly the dominant attribute for distinguishing between the hemiplegic affected leg patellar tendon reflex response and the unaffected leg patellar tendon reflex response. The research findings demonstrate the capacity of machine learning, such as the J48 decision tree, for attaining considerable classification accuracy that distinguishes between a hemiplegic affected patellar tendon Smartphone and Portable Media Device: A Novel Pathway toward the Diagnostic... http://dx.doi.org/10.5772/intechopen.69961 9

was attained through the amalgamation of a portable media device as a wireless accelerometer platform, a potential energy impact pendulum, and a machine learning algorithm. Each reflex response sample of the affected hemiplegic reflex response and unaffected reflex response was transmitted through local wireless connectivity as email attachments for post-processing. Software automation consolidated the trial data into a cohesive feature set. The selected machine learning algorithm was the support vector machine. This research demonstrates the considerable promise of the integration of smartphones and portable media devices as wearable and wireless inertial sensor platforms in combination with machine learning algorithms for clas-

Further investigation regarding the utility of the portable media device as a wireless accelerometer platform extended to the domain of assistive devices for gait, such as a cane. An issue with using a cane involves the concern that the patient is properly using the cane per the advice of a skilled therapist. However, such instruction is only realistically feasible respective of a relatively brief clinical appointment. A portable media device functioning as a wireless accelerometer platform can monitor the usage of a cane while the subject is walking, and the data package can be conveyed by local wireless connectivity to the Internet as an email attachment. Considerable classification accuracy was attained using logistic regression as a machine learning algorithm to distinguish between correct and incorrect cane usage scenarios with the

LeMoyne and Mastroianni demonstrated the potential of integrating a portable media device as a wireless gyroscope platform for quantifying the patellar tendon reflex response and machine learning for the classification of a hemiplegic reflex pair. The tendon reflex was elicited through supramaximal stimulation from a manually operated reflex hammer. The portable media device functioning as a wireless gyroscope platform was mounted proximal to

In order to consolidate the affected leg and unaffected leg reflex response into a feature set their respective gyroscope signals needed to be considered, such as presented in **Figure 2**. Three numeric attributes comprehensively characterize the nature of the reflex response:

• time disparity between maximum and minimum angular rate of rotation [45].

These three numeric attributes represent the reflex response in terms of both kinematic and temporal features. The feature set was evaluated through WEKA using the J48 decision tree. An advantage of the WEKA J48 decision tree is the ability to visualize the decision tree, which

The J48 decision tree is illustrated in **Figure3**. Note that the feature set attribute TimeMaxMinGamma is clearly the dominant attribute for distinguishing between the hemiplegic affected leg patellar tendon reflex response and the unaffected leg patellar tendon reflex response. The research findings demonstrate the capacity of machine learning, such as the J48 decision tree, for attaining considerable classification accuracy that distinguishes between a hemiplegic affected patellar tendon

sification of perceptively distinguishable human movement features [43].

recorded accelerometer signal deriving the feature set [44].

the lateral malleolus as demonstrated in **Figure 1** [45].

may infer the most predominant aspects of the feature set [45].

• maximum angular rate of rotation; • minimum angular rate of rotation;

8 Smartphones from an Applied Research Perspective

**Figure 1.** Portable media device as a wireless gyroscope platform for quantifying the patellar tendon reflex response [45].

reflex response and correlated unaffected patellar tendon reflex response through the quantification of a portable media device functioning as a wireless gyroscope platform [45].

An extension of the utility of using the portable media device as a wireless gyroscope platform with machine learning was applied to classify the usage of a proposed rehabilitation system. An ankle rehabilitation system intended to promote dorsiflexion was developed through 3D printing. The operation of the ankle rehabilitation system was quantified by mounting a smartphone functioning as wireless gyroscope platform to the foot plate of the ankle rehabilitation system [46].

**Figure 2.** Gyroscope signals for a hemiplegic reflex pair (affected leg (a) and unaffected leg (b)) for the patellar tendon reflex response [45].

**Figure 3.** J48 decision tree applied for attaining considerable classification accuracy for differentiating between a hemiplegic reflex pair for the patellar tendon reflex response [45].

The ability of the smartphone functioning as wireless gyroscope platform to convey its data sample as an email attachment through Internet connectivity makes the integral application suitable for homebound therapy scenarios. A subject could conduct a therapy session at home, and then the post-processing with machine learning classification could provide a therapist with advanced acuity of subject rehabilitation status. The gyroscope signal was distilled into a feature set for machine learning classification through a support vector machine, for which considerable classification accuracy was achieved [46].

Deep brain stimulation has been demonstrated as an efficacious treatment strategy for people with movement disorders that are intractable with medical therapy intervention. In particular essential tremor can be treated through the application of deep brain stimulation. However an issue is evident with respect to the considerable array of tuning parameters, which can present a daunting task for even an expert clinician. A smartphone as a wireless accelerometer platform presents a promising means for quantifying the efficacy of deep brain stimulation for people with essential tremor [47, 48].

**Figure 2.** Gyroscope signals for a hemiplegic reflex pair (affected leg (a) and unaffected leg (b)) for the patellar tendon

**Figure 3.** J48 decision tree applied for attaining considerable classification accuracy for differentiating between a

hemiplegic reflex pair for the patellar tendon reflex response [45].

reflex response [45].

10 Smartphones from an Applied Research Perspective

Preliminary research, development, testing, and evaluation of the capability to a smartphone as a wireless accelerometer platform to facilitate the tuning of a deep brain stimulator for ameliorating symptoms of essential tremor pertained to attaining machine learning classification accuracy for differentiating between the deep brain stimulator in 'On' and 'Off' states. Regarding both of these states a subject with essential tremor was tasked with reaching and grasping a lightweight object with a smartphone mounted to the dorsum of the hand. The trial data of the smartphone functioning as wireless accelerometer platform transmitted the recording as an email attachment with connectivity to the Internet. Subsequent post-processing of the acceleration signals derived a feature set that was applied to a machine learning algorithm. The most appropriate algorithm for the classification task under consideration based on the expertise of the research team was the support vector machine. The support vector machine achieved considerable classification accuracy for differentiating between a person with essential tremor conducting a simple reaching and grasping task with deep brain stimulation in 'On' and 'Off' status based on the acceleration signal of a smartphone functioning as a wireless accelerometer platform [47].

Reduced arm swing is a subject of concern for people with hemiplegic gait. Objectively quantifying the severity of reduced arm swing for the affected side contrasted to the unaffected side may enable a therapy intervention to ameliorate the impact of reduced arm swing during gait. Given the inherently rotational nature of arm swing during gait, this scenario is particularly relevant to the application of a smartphone functioning as a wireless gyroscope platform [49].

Especially regarding an outdoor environment for walking and evaluating reduced arm swing respective of hemiplegic gait, the smartphone functioning as a wireless gyroscope platform can readily transmit experimental data as an email attachment through its broad wireless connectivity to the Internet. The smartphone can be easily mounted proximal to the wrist joint through an armband as illustrated in **Figure 4**. Post-processing can be conveniently conducted in a remote location [49].

**Figure 4.** Mounting of the smartphone as a wireless gyroscope platform for quantifying reduced arm swing [49].

**Figure 5.** Gyroscope signals for reduced arm swing regarding the hemiplegic affected arm (a) and unaffected arm (b) [49].

**Figure 5** demonstrates the sample gyroscope signals of reduced arm swing of the affected arm contrasted to the unaffected arm. Based on the consideration of the gyroscope signal data the feature set was composed of aspects, such as descriptive statistics. A multilayer perceptron neural network was selected as the machine learning algorithm as provided in **Figure 6**. Considerable classification accuracy was attained for differentiating reduced arm swing regarding the hemiplegic affected arm and unaffected arm [49]. Similar results were also achieved with respect to another form of reduced arm swing manifested by Erb's Palsy [50].

Further refinement of the machine learning classification of a hemiplegic patellar tendon reflex pair was enabled through the application of the potential energy impact pendulum that provides predetermined amounts of potential energy to evoke the reflex and targeting of the reflex hammer. Upon observation of the obtained gyroscope signals the feature set was composed of three attributes:

Smartphone and Portable Media Device: A Novel Pathway toward the Diagnostic... http://dx.doi.org/10.5772/intechopen.69961 13

**Figure 6.** Multilayer perceptron neural network for classifying reduced arm swing regarding the hemiplegic affected arm and unaffected arm [49].

**Figure 7.** Gyroscope signals for the affected leg (a) and unaffected leg (b) regarding the patellar tendon reflex response [52].

• maximum angular rate of rotation;

**Figure 5** demonstrates the sample gyroscope signals of reduced arm swing of the affected arm contrasted to the unaffected arm. Based on the consideration of the gyroscope signal data the feature set was composed of aspects, such as descriptive statistics. A multilayer perceptron neural network was selected as the machine learning algorithm as provided in **Figure 6**. Considerable classification accuracy was attained for differentiating reduced arm swing regarding the hemiplegic affected arm and unaffected arm [49]. Similar results were also achieved with respect to another form of reduced arm swing manifested by Erb's Palsy [50]. Further refinement of the machine learning classification of a hemiplegic patellar tendon reflex pair was enabled through the application of the potential energy impact pendulum that provides predetermined amounts of potential energy to evoke the reflex and targeting of the reflex hammer. Upon observation of the obtained gyroscope signals the feature set was

**Figure 5.** Gyroscope signals for reduced arm swing regarding the hemiplegic affected arm (a) and unaffected arm (b) [49].

**Figure 4.** Mounting of the smartphone as a wireless gyroscope platform for quantifying reduced arm swing [49].

composed of three attributes:

12 Smartphones from an Applied Research Perspective


The multilayer perceptron neural network was selected as the machine learning algorithm, for which considerable classification accuracy was attained distinguishing between a hemiplegic patellar tendon reflex response and its associated unaffected patellar tendon reflex response [51].

A similar experiment was applied using supramaximal stimulation of the patellar tendon reflex through a manually operated reflex hammer. **Figure 7** illustrates the acquired gyroscope signals for the affected leg and unaffected leg regarding the patellar tendon reflex response. Subsequent post-processing derived the multilayer perceptron neural network presented in **Figure 8**. This research endeavor achieved appreciable classification accuracy for differentiating the hemiplegic (affected and unaffected) patellar tendon reflex pair [52].

Two recent developments regarding the role of smartphones and portable media device functioning as wireless gyroscope platforms for therapy and rehabilitation exercise have been applied to eccentric training using Virtual Proprioception and wobble board therapy. These two applications emphasize the flexibility of these devices, since the smartphone uses its broader telecommunication footprint regarding Virtual Proprioception for eccentric training in a gym setting and the more localized wireless connectivity of the portable media device in a homebound setting. Both experiments involve automated post-processing to develop a feature set for machine learning classification using a multilayer perceptron neural network [53, 54].

Virtual Proprioception for eccentric training is based on the successful application of Virtual Proprioception for real time gait rehabilitation. Virtual Proprioception for real time gait rehabilitation applied biofeedback based on a wireless accelerometer signals from a hemiplegic affected leg and unaffected leg while walking. Both auditory and visual feedback enabled the subject to modify gait in real time to achieve a closer state of parity regarding the acceleration waveforms of both legs. The wireless inertial sensors, such as the accelerometer, provide a virtual alternative to the neurological basis of proprioception [55, 56].

Eccentric strength training has be proposed as an highly effective means of strength training, however the rate of change is significant for the quality of the eccentric training event. Visual real time feedback from the gyroscope signal of a smartphone functioning as wireless gyroscope platform enables Virtual Proprioception for eccentric training. **Figure 9** illustrated the mounting of the smartphone through an armband proximal to the wrist joint for a biceps curl using Virtual Proprioception for eccentric training [54].

**Figure 8.** Multilayer perceptron neural network for classifying the affected leg and unaffected leg regarding the patellar tendon reflex response [52].

Smartphone and Portable Media Device: A Novel Pathway toward the Diagnostic... http://dx.doi.org/10.5772/intechopen.69961 15

Subsequent post-processing derived the multilayer perceptron neural network presented in **Figure 8**. This research endeavor achieved appreciable classification accuracy for differentiat-

Two recent developments regarding the role of smartphones and portable media device functioning as wireless gyroscope platforms for therapy and rehabilitation exercise have been applied to eccentric training using Virtual Proprioception and wobble board therapy. These two applications emphasize the flexibility of these devices, since the smartphone uses its broader telecommunication footprint regarding Virtual Proprioception for eccentric training in a gym setting and the more localized wireless connectivity of the portable media device in a homebound setting. Both experiments involve automated post-processing to develop a feature set for machine learning classification using a multilayer perceptron neural network

Virtual Proprioception for eccentric training is based on the successful application of Virtual Proprioception for real time gait rehabilitation. Virtual Proprioception for real time gait rehabilitation applied biofeedback based on a wireless accelerometer signals from a hemiplegic affected leg and unaffected leg while walking. Both auditory and visual feedback enabled the subject to modify gait in real time to achieve a closer state of parity regarding the acceleration waveforms of both legs. The wireless inertial sensors, such as the accelerometer, provide a

Eccentric strength training has be proposed as an highly effective means of strength training, however the rate of change is significant for the quality of the eccentric training event. Visual real time feedback from the gyroscope signal of a smartphone functioning as wireless gyroscope platform enables Virtual Proprioception for eccentric training. **Figure 9** illustrated the mounting of the smartphone through an armband proximal to the wrist joint for a biceps curl

**Figure 8.** Multilayer perceptron neural network for classifying the affected leg and unaffected leg regarding the patellar

ing the hemiplegic (affected and unaffected) patellar tendon reflex pair [52].

14 Smartphones from an Applied Research Perspective

virtual alternative to the neurological basis of proprioception [55, 56].

using Virtual Proprioception for eccentric training [54].

[53, 54].

tendon reflex response [52].

**Figure 9.** Virtual Proprioception for eccentric training demonstrated with a smartphone functioning as a wireless gyroscope platform secured proximal to the wrist through an armband [54].

During the eccentric muscle lengthening phase of a biceps curl the subject while using Virtual Proprioception for eccentric training uses the visual feedback from the smartphone gyroscope signal to ensure that a prescribed threshold is not exceeded. This inertial sensor acuity provides a virtual and quantified representation of neurologically derived proprioception. The control aspect of the experiment consists of an eccentric phase of a biceps curl without Virtual Proprioception. As illustrated in **Figure 10** these gyroscope signals display visually notable disparity [54].

The experimental trial data was transmitted wirelessly to the Internet as an email attachment. Automated post-processing of the data package developed a feature set based on the

**Figure 10.** Gyroscope signals for eccentric phase of a biceps curl respective of Virtual Proprioception for eccentric training (a) and without feedback (b) [54].

**Figure 11.** Multilayer perceptron neural network applied for classifying between scenarios of eccentric phase for a set of biceps curls respective of Virtual Proprioception for eccentric training and without feedback through Virtual Proprioception [54].

gyroscope signal data. Through the application of a multilayer perceptron neural network illustrated in **Figure 11** considerable classification accuracy was achieved for distinguishing between scenarios with Virtual Proprioception for eccentric training and eccentric training without feedback through Virtual Proprioception [54].

A wobble board is a therapy device that promotes rehabilitation of the ankle foot complex. A portable media device as shown in **Figure 12** can be readily mounted to the wobble board to provide advanced acuity as a wireless gyroscope platform. The rotation of the wobble board is notably disparate regarding the hemiplegic affected ankle and unaffected ankle upon observation. The observed disparity can be quantified through the wireless gyroscope platform from the portable media device and conveyed as an email attachment through wireless connectivity to the Internet. Post-processing consolidates the gyroscope signal into a feature set. With a multilayer perceptron neural network considerable classification accuracy was achieved for differentiating between the hemiplegic affected ankle and unaffected ankle [53].

Smartphone and Portable Media Device: A Novel Pathway toward the Diagnostic... http://dx.doi.org/10.5772/intechopen.69961 17

**Figure 12.** Wobble board for therapy of the ankle foot complex with a portable media device functioning as a wireless gyroscope platform [53].
