**The Emerging Wearable Solutions in mHealth**

Fang Zhao, Meng Li and Joe Z. Tsien

Additional information is available at the end of the chapter

http://dx.doi.org/10.5772/63557

#### **Abstract**

The marriage of wearable sensors and smartphones have fashioned a foundation for mobile health technologies that enable healthcare to be unimpeded by geographical boundaries. Sweeping efforts are under way to develop a wide variety of smartphonelinked wearable biometric sensors and systems. This chapter reviews recent progress in the field of wearable technologies with a focus on key solutions for fall detection and prevention, Parkinson's disease assessment and cardiac disease, blood pressure and blood glucose management. In particular, the smartphone-based systems, without any external wearables, are summarized and discussed.

**Keywords:** wearable inertial sensors, accelerometer, gyroscope, ECG patch, classifica‐ tion algorithm, smartphone, fall detection and prevention, Parkinson's disease, car‐ diac rhythm, blood glucose, blood pressure

#### **1. Introduction**

Nowadays, dramatic advances in microelectromechanical systems (MEMS) technology have paved the way for wearable sensors to make inroads into mHealth, providing the potential for medical care and research to take place outside the standard doctor's office or hospital. A wide variety of wearable biometric sensors, such as bracelets, watches, skin patches, headbands, earphones, and clothing [1, 2], have been designed and developed. Regardless of the various forms and functions of these sensors, their unifying design focus is to allow for unobtrusive, passive, and continuous monitoring. Beyond sensing capability, another key characteristic is their ability to seamlessly connect with a mobile device to transfer all biometric data into a software application (APP) that can be shared with healthcare providers, researchers or family members. Inertial sensors, the most ubiquitous wearables, combined with dedicated algo‐ rithms are able to "count" steps (i.e., pedometers), gauge physical activity levels, indirectly

© 2016 The Author(s). Licensee InTech. This chapter is distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/3.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

estimate energy expenditure [3], and implement activity recognition [4]. Today, the Holter monitor,themost commonlyusedambulatoryelectrocardiographydeviceforassessingcardiac abnormalities, is one of the technologies that may soon become obsolete, since prolonged continuous rhythm monitoring is available by wearing an electrocardiogram (ECG) patch on the chest [5]. Other notable examples of sensor technologies under development which allow for a more personalized understanding of our health include cuffless blood pressure monitor‐ ing and noninvasive blood glucose tracking. Through progressively miniaturized, smart‐ phones are equipped with comparatively advanced sensing capabilities (i.e., accelerometer, gyroscope, magnetometer, camera, and many more) and powerful computing capabilities, making it the ideal platform for remote health monitoring without the extra expense of purchasing and inconvenience of using dedicated wearables. As a result, smartphone-based solutions have emerged most recently for fall detection and prevention [6], activity recogni‐ tion[7],Parkinson'sdisease (PD) assessment[8], andcardiac rhythmmeasurementinmHealth.

This chapter provides a review of recent progress in the field of wearable systems and solutions that have already entered into or have the potential to apply in mHealth. Aging of the population is a global issue, and it presents tremendous challenges to society and healthcare systems all over the world. The most common healthcare issues of the aging population include the following: (i) falls that are considered as one of the major hazards for the elderly, especially for those living alone [9]; (ii) neurological disorders that are categorized as major chronic diseases inducing motor impairments, with PD as one of the most frequently occurring conditions [10]; and (iii) cardiac disease, hypertension and diabetes are the most common chronic diseases affecting the elderly [11]. Therefore, a critical analysis of the state-of-the-art wearable solutions for these age-related care issues and chronic diseases are presented.

The remainder of this chapter can be separated into five sections. The wearable solutions for motion monitoring are discussed in Section 2. Firstly, the basic architecture of the wearable motion monitoring systems is described, followed by a summary of the state-of-the-art smartphone-based fall detection and prevention systems, with a focus on the sensor used, extracted features, the classification algorithm, and the outcomes in each system. The wearable solutions for PD are then discussed. A selection of external wearable solutions and smart‐ phone-based systems that used pattern recognition algorithms to classify motor signs of functional activities impairment in PD are presented and compared. Section 3 illustrates the wearable solutions for cardiac activity monitoring. Several commercially available portable devices are presented. Section 4 describes the approaches for cuffless blood pressure moni‐ toring and noninvasive blood glucose monitoring. Unfortunately, these approaches are not satisfactory to date. Finally, conclusion offered in Section 5 points out important observations and areas that need further research.

#### **2. Wearable solutions for motion monitoring**

Mirroring the increasingly widespread adoption of wearable inertial sensors in personalized healthcare is an equally remarkable development in algorithms to classify human activity [7]. As a result, inertial sensor technologies can go well beyond step counts to a wealth of person‐ alized activity information to help guide health and wellness. Earlier work by Bouten *et al*. [12] established a significant relationship (*r*=0.89) between accelerometer output and energy expenditure due to physical activity, impelling wearable sensor to become capable of estimat‐ ing energy expenditure in diabetes or obesity management. Subsequent work by Najafi *et al*. [13] founded a significant correlation between postural transition (PT) and falling risk using a gyroscope, which led to a variety of other works to exemplify the prominence of wearable inertial sensors in fall detection and prevention in elderly care. The activity recognition by wearable inertial sensors has also been used in the assessment and rehabilitation of many neurological diseases [14], such as Parkinson's disease (PD), stroke, cerebral palsy (CP), multiple sclerosis (MS), and Huntington's disease (HD), which can induce motor impairment. Transformations are under way in movement monitoring to provide care in the daily lives of those afflicted with these diseases as a result of all these breakthroughs.

#### **2.1. Architecture**

estimate energy expenditure [3], and implement activity recognition [4]. Today, the Holter monitor,themost commonlyusedambulatoryelectrocardiographydeviceforassessingcardiac abnormalities, is one of the technologies that may soon become obsolete, since prolonged continuous rhythm monitoring is available by wearing an electrocardiogram (ECG) patch on the chest [5]. Other notable examples of sensor technologies under development which allow for a more personalized understanding of our health include cuffless blood pressure monitor‐ ing and noninvasive blood glucose tracking. Through progressively miniaturized, smart‐ phones are equipped with comparatively advanced sensing capabilities (i.e., accelerometer, gyroscope, magnetometer, camera, and many more) and powerful computing capabilities, making it the ideal platform for remote health monitoring without the extra expense of purchasing and inconvenience of using dedicated wearables. As a result, smartphone-based solutions have emerged most recently for fall detection and prevention [6], activity recogni‐ tion[7],Parkinson'sdisease (PD) assessment[8], andcardiac rhythmmeasurementinmHealth.

This chapter provides a review of recent progress in the field of wearable systems and solutions that have already entered into or have the potential to apply in mHealth. Aging of the population is a global issue, and it presents tremendous challenges to society and healthcare systems all over the world. The most common healthcare issues of the aging population include the following: (i) falls that are considered as one of the major hazards for the elderly, especially for those living alone [9]; (ii) neurological disorders that are categorized as major chronic diseases inducing motor impairments, with PD as one of the most frequently occurring conditions [10]; and (iii) cardiac disease, hypertension and diabetes are the most common chronic diseases affecting the elderly [11]. Therefore, a critical analysis of the state-of-the-art wearable solutions for these age-related care issues and chronic diseases are presented.

The remainder of this chapter can be separated into five sections. The wearable solutions for motion monitoring are discussed in Section 2. Firstly, the basic architecture of the wearable motion monitoring systems is described, followed by a summary of the state-of-the-art smartphone-based fall detection and prevention systems, with a focus on the sensor used, extracted features, the classification algorithm, and the outcomes in each system. The wearable solutions for PD are then discussed. A selection of external wearable solutions and smart‐ phone-based systems that used pattern recognition algorithms to classify motor signs of functional activities impairment in PD are presented and compared. Section 3 illustrates the wearable solutions for cardiac activity monitoring. Several commercially available portable devices are presented. Section 4 describes the approaches for cuffless blood pressure moni‐ toring and noninvasive blood glucose monitoring. Unfortunately, these approaches are not satisfactory to date. Finally, conclusion offered in Section 5 points out important observations

Mirroring the increasingly widespread adoption of wearable inertial sensors in personalized healthcare is an equally remarkable development in algorithms to classify human activity [7].

and areas that need further research.

4 Mobile Health Technologies - Theories and Applications

**2. Wearable solutions for motion monitoring**

The basic architecture of motion monitoring systems for mHealth consists of three common phases namely, sensing, processing and communication (**Figure 1**). Feature extraction and motion classification algorithm used in the processing phase may differ greatly from system to system.

**Figure 1.** Basic architecture of activity tracking systems for mHealth.

#### *2.1.1. Sensing*

Multimodal MEMS sensors can be utilized to identify physical activities, including acceler‐ ometer, gyroscope, magnetometer, barometer, etc. The terms accelerometer, gyroscope, and magnetometer will refer to triaxial accelerometers, triaxial gyroscopes and triaxial magneto‐ meters, respectively, unless otherwise stated. Each type of sensor is sensitive to a kinematic quantity: accelerometer for sensing acceleration along three orthogonal directions; gyroscope for detecting angular momentum; magnetometer for gauging changes in orientation by measuring the strength of the local magnetic field along three orthogonal axes; and barometer for determining rapid changes in altitude (e.g., walking up/down stairs) by measuring absolute atmospheric pressure to infer altitude above sea level. Their combination can even estimate three-dimensional (3D) orientation and displacement.

#### *2.1.2. Processing*

The processing phase encompasses preprocessing, feature extraction and physical motion classification steps. Preprocessing needs to be first applied to the raw data collected from MEMS sensors to improve the signal-to-noise ratio. The signals are often smoothed by median filters of a short sliding window to remove spurious noise [15]. Accelerometer data are often high-pass filtered to separate acceleration caused by gravity from acceleration due to body movement [16].

After preprocessing of MEMS data, features are generally extracted from sequential epochs of time using window techniques. The most commonly used approach is the sliding window often with 50% overlap between consecutive windows [17], which is the most suitable for realtime or online applications. Statistical measures of the time domain and frequency domain features are widely used to reduce the MEMS data of each window epoch to a finite number of derived parameters from which a physical movement can be inferred.

Prior to classification, feature selection techniques [18] may be applied to find the optimal feature subset, which can best distinguish between movements, from all of the features generated. Feature selection is of particular importance as inappropriate or redundant features may deteriorate the overall classification performance. The selected features from the MEMS sensor data are exploited by the classification algorithms in the development of a model that can identify specific physical movements. Classification methods used in activity recognition include (but are not limited to) hidden Markov models (HMM), K nearest neighbors (KNN), support vector machines (SVM), discrete wavelet transform (DWT), decision tree classifiers (DTC), random forests (RFs), linear discriminant analysis (LDA) or feed-forward neural network (Bpxnc).

#### *2.1.3. Communication*

After processing, the classified motion data can then be sent to medical staff (e.g., a caregiver or a physician) for remote monitoring or back to the user or patient for self-monitoring. Once an abnormal movement (i.e., fall event) is detected, the wearable mHealth systems sent out a signal to seek help from the monitoring center or a caregiver via smartphones.

#### **2.2. Fall detection and prevention**

Falls are one of the major causes of injuries and hospital admissions of elderly people. Those who suffer from neurological diseases (e.g., stroke, PD) also give rise to increased fall risks. Falls can potentially cause severe physical injuries, such as bleeding, fracture and central nervous system (CNS) damage, and long lie times (remaining involuntarily on the ground for a prolonged period) after the fall can lead to disability, paralysis, even death. Therefore, the first line of defence against fall hazards is to prevent them and the second line of defence is to provide emergency treatment in time.

#### *2.2.1. Smartphone-based systems*

atmospheric pressure to infer altitude above sea level. Their combination can even estimate

The processing phase encompasses preprocessing, feature extraction and physical motion classification steps. Preprocessing needs to be first applied to the raw data collected from MEMS sensors to improve the signal-to-noise ratio. The signals are often smoothed by median filters of a short sliding window to remove spurious noise [15]. Accelerometer data are often high-pass filtered to separate acceleration caused by gravity from acceleration due to body

After preprocessing of MEMS data, features are generally extracted from sequential epochs of time using window techniques. The most commonly used approach is the sliding window often with 50% overlap between consecutive windows [17], which is the most suitable for realtime or online applications. Statistical measures of the time domain and frequency domain features are widely used to reduce the MEMS data of each window epoch to a finite number

Prior to classification, feature selection techniques [18] may be applied to find the optimal feature subset, which can best distinguish between movements, from all of the features generated. Feature selection is of particular importance as inappropriate or redundant features may deteriorate the overall classification performance. The selected features from the MEMS sensor data are exploited by the classification algorithms in the development of a model that can identify specific physical movements. Classification methods used in activity recognition include (but are not limited to) hidden Markov models (HMM), K nearest neighbors (KNN), support vector machines (SVM), discrete wavelet transform (DWT), decision tree classifiers (DTC), random forests (RFs), linear discriminant analysis (LDA) or feed-forward neural

After processing, the classified motion data can then be sent to medical staff (e.g., a caregiver or a physician) for remote monitoring or back to the user or patient for self-monitoring. Once an abnormal movement (i.e., fall event) is detected, the wearable mHealth systems sent out a

Falls are one of the major causes of injuries and hospital admissions of elderly people. Those who suffer from neurological diseases (e.g., stroke, PD) also give rise to increased fall risks. Falls can potentially cause severe physical injuries, such as bleeding, fracture and central nervous system (CNS) damage, and long lie times (remaining involuntarily on the ground for a prolonged period) after the fall can lead to disability, paralysis, even death. Therefore, the first line of defence against fall hazards is to prevent them and the second line of defence is to

signal to seek help from the monitoring center or a caregiver via smartphones.

of derived parameters from which a physical movement can be inferred.

three-dimensional (3D) orientation and displacement.

6 Mobile Health Technologies - Theories and Applications

*2.1.2. Processing*

movement [16].

network (Bpxnc).

*2.1.3. Communication*

**2.2. Fall detection and prevention**

provide emergency treatment in time.

Initially, dedicated wearable kinematic sensors have been developed with the ability to assist in identifying falls [19, 20] and estimating the likelihood of future falls by monitoring activity levels or analyzing the individual's gait [21, 22]. However, their widespread adoption has been limited by the cost associated with purchasing the device and the low utilization coefficient by the user (who may often forget or refuse to wear the specially designed wearables). There has been a shift toward smartphones in recent years, as the smartphone with multimodal builtin MEMS sensors, coupled with its ubiquitous nature and increased computational power, make it the ideal platform for fall monitoring in mHealth. The first smartphone-based fall detection app iFall [23] utilized an integrated accelerometer to recognize the difference in position before and after the fall. Later in 2010, the PreFallD [24] was developed considering both the wearer's acceleration and orientation during the fall event. **Table 1** summarizes and compares the features of the existing smartphone-based fall detection and prevention systems or applications. The literatures that presented very preliminary investigations and did not declare the performance of their proposed solutions are not included here.




**Table 1.** Smartphone-based fall detection and prevention systems.

**Arti cle** 

**Appli cation**  **Sensors (Placement)**

8 Mobile Health Technologies - Theories and Applications

gyroscope

pocket)

(waist)

(waist)

Accelerometer & gyroscope (waist)

(shirt, or trouser

(trouser pocket)

gyroscope (hand, pocket, waist)

(chest, waist, thigh)

gyroscope (hand,

& Wi-Fi module

pocket)

(waist)

pocket)

[32] Detection Accelerometer (Shirt pocket)

(hand, shirt, or trouser

Threshold, One-class SVM

[39] Detection Accelerometer &

[98] Detection Accelerometer

[44] Detection Accelerometer

[99] Detection Accelerometer

[40] Detection Accelerometer

[27] Detection Accelerometer &

[33] Detection Accelerometer

[34] Detection Accelerometer &

[41] Detection Accelerometer

[6] Detection, prevention **Algorithm Performance Notification**

77.9412% (*AC* for shirt pocket);

75% (*AC* for hand);

between running and

on a smartphone to realize long-term and real-time

threshold 92.75% (*SE*), 86.75% (*SP*) Text message

SVM 95.7% (*PR*), 90% (average *RC*) vibration, audio

Threshold 72.22% (*SE*), 73.78 (*SP*) SMS

Threshold 80% (*SE*), 96.25% (*SP*), 85% (*AC*) Undisclosed

100% & 75.8% (*PR* & *RC* for DT); 99.81% & 75.43% (*PR* & *RC* for SVM); 98.67% & 73.20%

85.3% (*SE*), 90.5% (*SP*) Undisclosed

84.2857% (*AC* for trouser pocket)

Threshold Capability of differentiate

falling

Threshold, ANN 100% success rate for a total of 500 epochs.

Threshold The uFall and uTUG can ran

monitoring.

(average *SP*)

Threshold 97% (average *SE*), 100%

Semisupervised learning

DT, SVM, NB,

RSSI

**(Information)**

Undisclosed

SMS (time, GPS coordinates)

Message (GPS coordinates)

SMS.

Undisclosed

alarm, SMS (time,

SMS (name, time,

location)

location)

Audio alarm, email,

The most common sensor used in fall detection and prevention was the accelerometer, followed by the gyroscope (**Table 1**). In most of the studies, threshold-based algorithm was adopted for fall detection due to its low complexity. The most commonly used feature for threshold-based algorithm is the magnitude vector of acceleration signal:

$$<\langle A\_T \vert = \sqrt{\left| \left| A\_{\mathbf{x}} \right|^2 + \left| \left| A\_{\mathbf{y}} \right|^2 + \left| \left| A\_{\mathbf{z}} \right|^2 \right| \right.}\tag{1}$$

where *Ax*, *Ay*ḥ, and *Az* represent accelerometer signals of the *x-, y-*, and *z*-axis, respectively. The threshold value could be predefined (fixed) or adaptive (changed with user-provided phys‐ iological data, such as height, weight).

The surge in computing power has fashioned a foundation for complex machine-learning classification algorithms for fall detection and prevention to be implemented in smartphones. The classification algorithms used in the processing phase vary considerably across systems. Yavuz *et al*. [25] utilized DWT and achieved a better true-positive (TP) performance while decreasing the false positives (FP) when compared to threshold-based algorithm. Zhao *et al*. [26] implemented three machine-learning algorithms—namely C4.5 DTC, NB, and SVM and compared their performances based on recognition accuracy. Fahmi *et al*. [27] designed a semisupervised algorithm to detect a genuine fall event with smartphone. He and Li [28] employed a combined algorithm of Fisher's discriminant ratio (FDR) criterion and J3 criterion for feature selection and hierarchical classifiers to recognize 15 activities including fall events. Majumder *et al*. [29] applied Hjorth mobility and complexity to identify high-risk gait patterns, hence developed a fall prevention system called iPrevention.

Once a fall event is detected, the systems send out notifications including audible alarms, vibrations, automatic voice calls, short message service (SMS), multimedia messaging service (MMS), E-mails, Twitter messaging, etc., (**Table 1**). Notification messages may contain infor‐ mation regarding time and location (GPS coordinates or Google Map).

#### *2.2.2. Performance evaluations*

There is no uniform standard for outcome evaluations of fall detection or prevention systems now. The outcomes are often represented by four possible situations [24, 30]: *TP*, a fall occurred and was correctly detected; *FP*, the system declared a fall that did not occur; true negative (TN), a fall-like event was not misclassified as a fall event; false negative (FN), a fall occurred, but the system missed it. The reliability of systems is usually evaluated based on the following parameters: sensitivity (*SE*) = *TP*/(*TP*+*FN*), which is the ratio of fallers correctly classified as fall event [27, 31–34]; specificity (*SP*) = *TN*/(*TP*+*FN*), which is the ratio of fall-like events correctly classified as nonfallers [35–38]; accuracy = (*TP+TN*)/(*TP*+FP+*FN+TN*), which is the ratio of true results in the whole data set [26, 28, 29, 39]. Some works measured the performance in a different way; they utilized *precision* = (∩)/and recall—namely, the number of correct results divided by the total outputs—as the performance indexes [40–42]. Some other works evaluated the proposed system by measuring the area under the receiver operating characteristic curve (AUC), where the curve represented *SE* versus *FN* [43].

#### *2.2.3. Limitations and challenges*

The most common sensor used in fall detection and prevention was the accelerometer, followed by the gyroscope (**Table 1**). In most of the studies, threshold-based algorithm was adopted for fall detection due to its low complexity. The most commonly used feature for

where *Ax*, *Ay*ḥ, and *Az* represent accelerometer signals of the *x-, y-*, and *z*-axis, respectively. The threshold value could be predefined (fixed) or adaptive (changed with user-provided phys‐

The surge in computing power has fashioned a foundation for complex machine-learning classification algorithms for fall detection and prevention to be implemented in smartphones. The classification algorithms used in the processing phase vary considerably across systems. Yavuz *et al*. [25] utilized DWT and achieved a better true-positive (TP) performance while decreasing the false positives (FP) when compared to threshold-based algorithm. Zhao *et al*. [26] implemented three machine-learning algorithms—namely C4.5 DTC, NB, and SVM and compared their performances based on recognition accuracy. Fahmi *et al*. [27] designed a semisupervised algorithm to detect a genuine fall event with smartphone. He and Li [28] employed a combined algorithm of Fisher's discriminant ratio (FDR) criterion and J3 criterion for feature selection and hierarchical classifiers to recognize 15 activities including fall events. Majumder *et al*. [29] applied Hjorth mobility and complexity to identify high-risk gait patterns,

Once a fall event is detected, the systems send out notifications including audible alarms, vibrations, automatic voice calls, short message service (SMS), multimedia messaging service (MMS), E-mails, Twitter messaging, etc., (**Table 1**). Notification messages may contain infor‐

There is no uniform standard for outcome evaluations of fall detection or prevention systems now. The outcomes are often represented by four possible situations [24, 30]: *TP*, a fall occurred and was correctly detected; *FP*, the system declared a fall that did not occur; true negative (TN), a fall-like event was not misclassified as a fall event; false negative (FN), a fall occurred, but the system missed it. The reliability of systems is usually evaluated based on the following parameters: sensitivity (*SE*) = *TP*/(*TP*+*FN*), which is the ratio of fallers correctly classified as fall event [27, 31–34]; specificity (*SP*) = *TN*/(*TP*+*FN*), which is the ratio of fall-like events correctly classified as nonfallers [35–38]; accuracy = (*TP+TN*)/(*TP*+FP+*FN+TN*), which is the ratio of true results in the whole data set [26, 28, 29, 39]. Some works measured the performance in a different way; they utilized *precision* = (∩)/and recall—namely, the number of correct results divided by the total outputs—as the performance indexes [40–42]. Some other works evaluated the proposed system by measuring the area under the receiver operating characteristic curve

<sup>222</sup> | |||| ||| *AAAA Txyz* = ++ (1)

threshold-based algorithm is the magnitude vector of acceleration signal:

hence developed a fall prevention system called iPrevention.

(AUC), where the curve represented *SE* versus *FN* [43].

mation regarding time and location (GPS coordinates or Google Map).

iological data, such as height, weight).

10 Mobile Health Technologies - Theories and Applications

*2.2.2. Performance evaluations*

Despite the expanding body of evidence to support the use of smartphones for fall detection and prevention, it is important to recognize the limitations in this area of science. The promi‐ nent weakness is problems induced by the limited battery life of the smartphone. The rate at which the smartphone's battery is consumed is dependent on both internal and external factors. Internal factors are built-in sensor dependent, including the sampling rate and resolution mode. High-resolution mode can dramatically increase the rate of power consump‐ tion. External factors are related to the number of sensors used, data recording time, and complexity of the algorithms. Mellone *et al*. [6] showed that a battery could power a smart‐ phone (Samsung Galaxy S II) for 30 h with only one sensor used and 16 h with three sensors activated. Majumder *et al*. [29] reported that a fully charged battery can only power an iPhone for 3 h at the most, when running a machine learning algorithm. Energy efficiency will continue to be an important criterion when choosing the algorithm, unless advancements in battery technology could lead to higher density energy storage.


**Table 2.** Specifications of the built-in sensors in some currently available smartphones.

The resolution and dynamic range of the built-in inertial sensors vary considerably across smartphones (**Table 2**). Acceptable dynamic ranges for accelerometers from ±4 g to ± 16 g (*g* = 9.81 ms−2) have been reported for fall detection applications [35, 44], which is beyond the typical dynamic ranges of most currently available smartphone accelerometers ( ± 2 g). The newest high-end commercially available smartphones (i.e., iPhone 6/6plus) have accelerome‐ ters with higher dynamic ranges ( ± 8 g), making these devices more suitable for detecting falls.

In addition, a major limitation of using smartphones to detect fall is that it requires the smartphone to be consistently located and/or oriented in the same position. It may be difficult to do so due to the multifunctional nature of smartphones. Habib *et al*. [45] showed that individuals may not place their smartphone on their body whilst at home so, that being said, it may limit the ability of the smartphone to detect fall in the home. At present, smartphone placement and usability issues should be handled carefully.

#### **2.3. Functional activities assessment for Parkinson' s disease**

For a population that is shifting toward an older age range, PD is categorized in the most common chronic neurological disorders. PD is characterized as an age-related neurodegener‐ ative disorder due to the loss of dopamine-producing brain neurons, an important neuro‐ transmitter involved in the regulation of movement. Progressive tremor, bradykinesia, hypokinesia, rigidity, and impaired postural control are common and disabling features of most patients with PD. The motor disorder analysis is generally performed in a clinical setting to provide subjective assessments. However, the motor fluctuation measurements in the clinical setting might not precisely reveal the real functional disability experienced by patients in natural environment. With the existing and on-going advance developments in MEMS technologies, continuous, unsupervised, objective and reliable monitoring of mobility and functional activities in natural environments is now possible, allowing for long-term, homebased intensive care and improvement of the individual healthcare and well being.

#### *2.3.1. Wearable inertial sensor-based methods*

A growing body of literature studied the use of wearable inertial sensors to detect and quantify tremor, bradykinesia and levodopa-induced dyskinesia (LID) in PD populations. Most studies were focused on finding the features derived from sensor signals that are effective for detecting differences between people with PD and healthy controls [46–49]. Results from these studies presented a range of outcomes which included the root mean square (RMS) of accelerations, the deviation of acceleration, step or stride variability, gait regularity or symmetry, FFT features, entropy and many more. Only a few works established and validated motion analysis methods or systems that used pattern recognition algorithms to classify motor signs of functional activities impairment in PD. **Table 3** provides a detailed comparison of these different methodological approaches. Leave-one-subject-out method and cross-validation method were used for validating the approaches.


The resolution and dynamic range of the built-in inertial sensors vary considerably across smartphones (**Table 2**). Acceptable dynamic ranges for accelerometers from ±4 g to ± 16 g (*g* = 9.81 ms−2) have been reported for fall detection applications [35, 44], which is beyond the typical dynamic ranges of most currently available smartphone accelerometers ( ± 2 g). The newest high-end commercially available smartphones (i.e., iPhone 6/6plus) have accelerome‐ ters with higher dynamic ranges ( ± 8 g), making these devices more suitable for detecting falls.

In addition, a major limitation of using smartphones to detect fall is that it requires the smartphone to be consistently located and/or oriented in the same position. It may be difficult to do so due to the multifunctional nature of smartphones. Habib *et al*. [45] showed that individuals may not place their smartphone on their body whilst at home so, that being said, it may limit the ability of the smartphone to detect fall in the home. At present, smartphone

For a population that is shifting toward an older age range, PD is categorized in the most common chronic neurological disorders. PD is characterized as an age-related neurodegener‐ ative disorder due to the loss of dopamine-producing brain neurons, an important neuro‐ transmitter involved in the regulation of movement. Progressive tremor, bradykinesia, hypokinesia, rigidity, and impaired postural control are common and disabling features of most patients with PD. The motor disorder analysis is generally performed in a clinical setting to provide subjective assessments. However, the motor fluctuation measurements in the clinical setting might not precisely reveal the real functional disability experienced by patients in natural environment. With the existing and on-going advance developments in MEMS technologies, continuous, unsupervised, objective and reliable monitoring of mobility and functional activities in natural environments is now possible, allowing for long-term, home-

based intensive care and improvement of the individual healthcare and well being.

A growing body of literature studied the use of wearable inertial sensors to detect and quantify tremor, bradykinesia and levodopa-induced dyskinesia (LID) in PD populations. Most studies were focused on finding the features derived from sensor signals that are effective for detecting differences between people with PD and healthy controls [46–49]. Results from these studies presented a range of outcomes which included the root mean square (RMS) of accelerations, the deviation of acceleration, step or stride variability, gait regularity or symmetry, FFT features, entropy and many more. Only a few works established and validated motion analysis methods or systems that used pattern recognition algorithms to classify motor signs of functional activities impairment in PD. **Table 3** provides a detailed comparison of these different methodological approaches. Leave-one-subject-out method and cross-validation

placement and usability issues should be handled carefully.

12 Mobile Health Technologies - Theories and Applications

**2.3. Functional activities assessment for Parkinson' s disease**

*2.3.1. Wearable inertial sensor-based methods*

method were used for validating the approaches.


2 AAM-active arm movement.

**Table 3.** Wearable inertial sensor-based methods for Parkinson's disease.

These methods were founded on various machine-learning classifiers. Salarian *et al*. [50] applied a fuzzy classifier combined with a logistic regression model to categorize sit-to-stand (STS) transitions. Three inertial sensors were used to detect the kinematic features of the trunk movements during the transitions. Compared to video recordings reference system, it demonstrated the ability to differentiate sit-to-stand from stand-to-sit with a sensitivity of 83.3% in PD and 94.4% in controls. Another study by Mazilu *et al*. [51] presented the GaitAssist system to detect FoG with two ankle-mounted IMUs, streaming data via Bluetooth to an Android phone. Supervised machine-learning models, trained offline using several FFT features, were utilized with an overall FoG hit rate of 94.94% and a specificity of 94%.

Some studies, on the other hand, evaluated various classifiers to identify ambulatory activities. Cancela *et al*. [52] implemented six activity recognition algorithms, —namely KNN, Parzen, Parzen density, DTC, Bpxnc, and SVM, to detect the severity of bradykinesia and found out that the SVM revealed the best classification results with 86% sensitivity by using two features (RMS and range). Barth *et al*. [53] employed three classifiers, including boosting with decision stump, LDA and SVM, to measure gait patterns in PD to distinguish mild and severe gait impairment. The system was able to classify PDs and controls with 88% sensitivity and 86% specificity using the LDA classifier based on three activities—namely 10 m walking, heel-toe tapping, and foot circling. It reached a 100% sensitivity and specificity to distinguish mild from severe using optimal features—namely step duration, entropy, variance, energy ratio, and a 0.5–3 Hz energy band. Klucken *et al*. [54] used 694 features and three pattern recognition algorithms (LDA, AdaBoost, and SVM) to categorize patients in different stages. The devel‐ oped eGaIT system, which consists of accelerometers and gyroscopes attached to shoes, was able to successfully distinguish patients from controls with an overall classification rate of 81%. The classification accuracy increased to 91% for more severe motor impairment or H&Y III patients.

Besides evaluating classifier, other works provided a complete motor assessment by analyzing the severity of several PD motor symptoms. Zwartjes *et al*. [55] used DTC to analyze motor activity and the severity of tremor, bradykinesia, and hypokinesia in patients with PD at three different levels of deep brain stimulation (DBS) treatment. An overall accuracy of 99.3% was achieved. Tzallas *et al*. [56] developed a system called PERFORM, using four accelerometers at each extremity and one accelerometer/gyroscope on the waist, to evaluate and quantify various symptom severity. The severity and type of tremor were classified by HMM classifier based on several time and frequency domain characteristics with 87% accuracy and 0.008 mean absolute error (*MBE*). The C4.5 DCT algorithm was used for LID detection and severity classification with an accuracy of 85.4% and a *MBE* of 0.31. A SVM classifier with optimum features (including approximate entropy, across correlation value, and range value) achieved 74.5% accuracy and 0.25 *MBE* for bradykinesia assessment. The detection of FoG was realized by an RFs classifier using the boot strap technique with 79% accuracy and 0.79 *MBE*. The PERFORM system also included a local base unit and a centralized hospital unit, allowing for the continuous remote monitoring and management of patients with PD.

#### *2.3.2. Smartphone-based solutions*

**Arti cle** 

56 Four

1

2

accelerometers (extremity) and one accelerometers & gyroscopes (waist)

AAM-active arm movement.

HMM (for tremor)

14 Mobile Health Technologies - Theories and Applications

SVM (for Bradykinesia)

IAA-Integrals of the absolute value of the accelerometer output;

**Table 3.** Wearable inertial sensor-based methods for Parkinson's disease.

**Sensors (Placement)**  **Algorithm Features Performance Validity**

method DT (for LID) Mean value, standard

87% (*AC*), 0.008 (*MBE*)

85.4% (*AC*), 0.31 (*MBE*)

74.5% (*AC*), 0.25 (*MBE*)

0.79 (*MBE*)

Leave-onesubject-out

on 0.5–3 Hz, *Pfreeze* on 3–8

Time, frequency and spatial features.

deviation, entropy, energy in specific frequency subbands,

Approximate entropy, sample entropy, RMS, cross correlation, range.

These methods were founded on various machine-learning classifiers. Salarian *et al*. [50] applied a fuzzy classifier combined with a logistic regression model to categorize sit-to-stand (STS) transitions. Three inertial sensors were used to detect the kinematic features of the trunk movements during the transitions. Compared to video recordings reference system, it demonstrated the ability to differentiate sit-to-stand from stand-to-sit with a sensitivity of 83.3% in PD and 94.4% in controls. Another study by Mazilu *et al*. [51] presented the GaitAssist system to detect FoG with two ankle-mounted IMUs, streaming data via Bluetooth to an Android phone. Supervised machine-learning models, trained offline using several FFT

features, were utilized with an overall FoG hit rate of 94.94% and a specificity of 94%.

Some studies, on the other hand, evaluated various classifiers to identify ambulatory activities. Cancela *et al*. [52] implemented six activity recognition algorithms, —namely KNN, Parzen, Parzen density, DTC, Bpxnc, and SVM, to detect the severity of bradykinesia and found out that the SVM revealed the best classification results with 86% sensitivity by using two features (RMS and range). Barth *et al*. [53] employed three classifiers, including boosting with decision stump, LDA and SVM, to measure gait patterns in PD to distinguish mild and severe gait impairment. The system was able to classify PDs and controls with 88% sensitivity and 86% specificity using the LDA classifier based on three activities—namely 10 m walking, heel-toe tapping, and foot circling. It reached a 100% sensitivity and specificity to distinguish mild from severe using optimal features—namely step duration, entropy, variance, energy ratio, and a

RF (for FOG) Entropy 79% (*AC*),

entropy.

Hz,

freeze index.

Given that smartphones are ubiquitous and have advanced built-in inertial sensors, research has recently sought to develop smartphone-based systems for PD assessments, which can keep the patient "connected" to his physician on a daily basis. The important features of existing smartphone-based solutions are summarized and compared in **Table 4**.



DFA-Extent of randomness; 4 CV-Coefficient of variation.

**Table 4.** Smartphone-based solutions for Parkinson's disease.

These smartphone-based solutions use the signal from the integrated accelerometers or gyroscopes in consumer-grade smartphones and in conjunction with machine learning algorithms to quantify key movement severity symptoms (i.e., bradykinesia, FoG, hand tremor) and discriminate patients with PD from controls. Arora *et al*. [57] using an RFs classifier with a range of different time and frequency features of the acceleration time series, achieved 98.5% average sensitivity and 97.5% average specificity in differentiating patients with PD from controls. Another study by Printy *et al*. [8] developed an iPhone application using embedded hardware of a smartphone, including gyroscope, accelerometer, capacitive touch screen, microphone, and the front-facing camera, and a SVM algorithm to discriminate between more severe and less severe bradykinesia with an accuracy of 94.5%. The accurate classification of bradykinesia severity was not achieved in this work.

Some studies, on the other hand, aimed to detect FoG, a common motor impairment to suffer an inability to walk in PD patients. Pepa *et al*. [58] presented a smartphone-integrated accel‐ erometer-based system to detect the FoG. They developed a linguistic fuzzy modelling (LFM) with Mamdani rule structure by fusing the information of freeze index, energy sum, cadency variation, and energy derivative ratio with a sensitivity of 89% and a specificity of 97%. In the smartphone-based system for FoG detection proposed by Kim *et al*. [59], data are derived from both embedded accelerometer and gyroscope. An AdaBoost.M1 classifier using several time and frequency domain features showed the best sensitivity of 86% at the waist, 84% and 81% in the trouser pocket and at the ankle, respectively.

Two other studies used the smartphone to measure the hand tremor symptom. Kostikis *et al*. [60] utilized a Breiman's RFs to classify upper limb tremor and achieved 82% accuracy in patients with PD and 90% accuracy in controls, with 0.9435 AUC. The feature metrics were derived from the acceleration vector and rotational velocity vector when patients performed two MDS-UPDRS postures—namely "Extended" and "Rest". Pan *et al*. [61] designed a prototype mobile cloud-based mHealth app on the Android platform called "PD Dr" to measure the severity of both hand resting tremor and gait difficulty, using the built-in accelerometer. The SVM classifier was used with a sensitivity of 77% and a specificity of 82% for hand resting tremor detection, and 89% sensitivity and 81% specificity in gait difficulty detection. Lasso regression approach was built to estimate the symptom severity. There was a strong correlation with PD disease stage (*r*=0.81), hand resting tremor severity(*r*=0.74), and gait difficulty severity (*r*=0.79).

#### *2.3.3. Limitations and challenges*

**Arti cle** 

**Sensors (Placement)** 

8 Accelerometer & gyroscope & touch screen & microphone

58 Accelerometer (hips)

60 Accelerometer & gyroscope (hand)

59 Accelerometer & gyroscope (ankle, trouser pocket, waist, chest pocket)

61 Accelerometer (hand or ankle)

SD-Standard Deviation;

IQR-Inter-quartile range;

DFA-Extent of randomness;

CV-Coefficient of variation.

1

2

3

4

Average frequencies, RMS angular velocity, speed of movement, amplitude of dominant

, PSD,

cadency variation, the ratio of the

Magnitude of acceleration and rotational velocity, SD of acceleration, mean magnitude of rotation rate.

Mean, variance, SD, entropy,

Hand tremor: power between 4–6 Hz, fraction of power, power ration in 3.5–15 Hz to 0.15– 3.5 Hz, total power from 0– 20 Hz, peak power, average acceleration.Gait: average gait cycle, average stride length, average walking speed, average acceleration, the

number of steps and the speed

These smartphone-based solutions use the signal from the integrated accelerometers or gyroscopes in consumer-grade smartphones and in conjunction with machine learning

of turning 360°.

**Table 4.** Smartphone-based solutions for Parkinson's disease.

energy, Fi, power, RMS, interquantile range, kurtosis, frequency domain

rhythm, CV4

16 Mobile Health Technologies - Theories and Applications

RMS values.

derivative of the energy.

features.

Freeze index, energy,

**Features Algorithm Performance Validity**

fuzzy Logic algorithm

AdaBoost. M1,

SVM 77% & 82%

gait

SVM, RFs 94.5% (*AC*), >

0.85 (AUC)

BagDT 82% (*AC* in patients),

90% (*AC* in controls)

86% & 84% & 81% (*SE* at the waist, in the trouser pocket and at the ankle, respectively).

(*SE* & *AC* for hand resting tremor detection), 89% & 81% (*SE* & *AC* for

difficulty detection).

Cross validation

Cross validation

Cross validation

Crossvalidation

89% (*SE*), 97% (*SP*) Undisclosed

Given the relatively small number of classifier-based studies in this area and the wide variety of research questions addressed, ranging from activity classification to different symptom severity level assessment, it is currently difficult to address which classifier is ideal in PD populations for mHealth. Meanwhile, the accuracy levels of the classifiers were generalized on small sample sizes ranging from 5 to 27 subjects [50–53, 55–61]. Only one out of these studies enlisted a relatively larger sample of 92 patients with PD and 81 controls [54]. It is therefore important to evaluate the performance of classifiers according to larger, homogeneous population sets. Moreover, It is difficult to evaluate how effective or well performing of a classifier, because its performance also depends on the selected features and the properties of wearable sensors (i.e., resolution, noise level). Therefore, the effectiveness of wearable inertialbased methods in mHealth regimens still has to be further examined.

Using a smartphone for PD management seems promising in mHealth, yet there are the same issues as those in smartphone-based fall detection systems. The performance and usability of smartphone-based solutions remain limited by the relatively lower quality of embedded sensors, and the limited battery life of smartphones, as well as the need to wear the smartphone in a fixed position.

Only very few studies provided a complete overall assessment of PD [55, 56]. Most of the existing solutions with external wearables sensors or the smartphones built-in sensors have limited focus on a particular motor symptom, and lack the important characteristic for PDmonitoring services, such as long-term recording, qualitative and quantitative assessments. Therefore, more effort should be put into providing a complete tool that comprises the most common PD motor disabilities, such as tremor, bradykinesia, LID, and FoG.
