© Schattauer 2016 Methods Inf Med 6/2016

1

Wea

rabl

e Th

erap

y

Focus Theme – Original Articles

Unobtrusive and Continuous Monitoring of Alcohol-impaired Gait Using Smart ShoesEunjeong Park1; Sunghoon I. Lee2; Hyo Suk Nam3; Jordan H. Garst4; Alex Huang4; Andrew Campion4; Monica Arnell4; Nima Ghalehsariand4; Sangsoo Park5; Hyuk-jae Chang6; Daniel C. Lu4; Majid Sarrafzadeh7

1Cardiovascular Research Institute, Yonsei University College of Medicine, Seoul, Korea; 2Department of Physical Medicine & Rehabilitation, Harvard Medical School, Boston, Massachusetts, USA; 3Department of Neurology, Yonsei University College of Medicine, Seoul, Korea; 4Department of Neurosurgery, University of California in Los Angeles (UCLA), Los Angeles, California, USA; 5Department of Computer Science and Engineering, Ewha Womans University, Seoul, Korea; 6Department of Cardiology, Yonsei University College of Medicine, Seoul, Korea; 7Computer Science Department, University of California in Los Angeles (UCLA), Los Angeles, California, USA

KeywordsWearable devices, personal monitoring, gait analysis, alcohol monitoring, wireless health-care

SummaryBackground: Alcohol ingestion influences sensory-motor function and the overall well-being of individuals. Detecting alcohol- induced impairments in gait in daily life necessitates a continuous and unobtrusive gait monitoring system.Objectives: This paper introduces the devel-opment and use of a non-intrusive monitor-ing system to detect changes in gait induced by alcohol intoxication.Methods: The proposed system employed a pair of sensorized smart shoes that are equipped with pressure sensors on the in-sole. Gait features were extracted and ad-justed based on individual’s gait profile. The adjusted gait features were used to train a

machine learning classifier to discriminate alcohol-impaired gait from normal walking. In experiment of pilot study, twenty partici-pants completed walking trials on a 12 meter walkway to measure their sober walking and alcohol-impaired walking using smart shoes.Results: The proposed system can detect al-cohol-impaired gait with an accuracy of 86.2 % when pressure value analysis and per-son-dependent model for the classifier are applied, while statistical analysis revealed that no single feature was discriminative for the detection of gait impairment.Conclusions: Alcohol-induced gait disturb-ances can be detected with smart shoe technology for an automated monitoring in ubiquitous environment. We demonstrated that personal monitoring and machine learn-ing-based prediction could be customized to detect individual variation rather than apply-ing uniform boundary parameters of gait.

Correspondence to:Eunjeong Park, PhD Cardiovascular Research InstituteYonsei University College of Medicine50 –1 Yonsei-roSeodaemun-gu, Seoul, 03722KoreaE-mail: eunjeong-park@yuhs.ac

Methods Inf Med 2016;55: ■–■http://dx.doi.org/10.3414/ME15-02-0008received: October 26, 2015accepted in revised form: July 27, 2016epub ahead of print: October 26, 2016

FundingThis study was supported by grants from the R&D Pro-gram of Fire Fighting Safety and 119 Rescue Technol-ogy , Ministry of Public Safety and Security, Republic of Korea (MPSS -2015 -70).

1. IntroductionExcessive alcohol use has been associated to 4 % of all death worldwide [1] and the fourth leading preventable cause of death in the United States [2, 3]. Even mild consumption of alcohol has been linked to adverse health effects, and therefore, pregnant mothers or patients with hepatic diseases are advised to maintain complete abstinence [4 – 6]. The rate of motor vehicle accidents, falls, as-saults, homicides, and suicides has also been reported to be higher with regular alcohol consumption [7– 9]. Unintentional injuries related to alcohol consumption are also identified as a major source of immediate health risks [10, 11]. Therefore, recognition of alcohol-induced effects on individuals could contribute significantly to lowering the risk for injury, and enables the medical treatment to be customized into each pa-tient’s pattern of alcohol consumption. Es-pecially, sensor-equipped monitoring tools provide accurate information which indi-viduals intentionally miss to report, or are hard to record [12].

The alcohol sensors had been categorized to detect blood alcohol content (BAC) and transdermal alcohol content (TAC). Breath-alyzer is a widely used tool that directly measures BAC. However, breathalyzers en-able relatively sporadic monitoring and require user’s voluntary meas urement rather than automated monitoring. As a mobile sensing tool of TAC, the SCRAM (Secure Continuous Alcohol Monitoring) system can determine alcohol levels expelled through the

For personal or educational use only. No other uses without permission. All rights reserved.Note: Uncorrected proof, epub ahead of print online

Downloaded from www.methods-online.com on 2016-10-28 | IP: 64.60.11.34

Methods Inf Med 6/2016 © Schattauer 2016

2

Wea

rabl

e Th

erap

y

E. Park et al.: Alcohol-impaired Gait Monitoring Using Smart Shoes

perspiration from skin, using ankle bracelets [13]. However, the weight of the ankle brace-let may restrict an individual’s movements and its direct contact to the skin may cause ir-ritation, which influences users’ long-term adherence to monitoring. Moreover, the type of ankle bracelet may reduce user’s adherence to the device, since ankle monitors are also used as a deterrent to criminal behavior. As addressed in [14], wearables which are de-fined as electronic devices that can be em-bedded in the user’s outfit, should not prevent the user’s activities. In this respect, a wearable device for monitoring of alcohol-induced im-pairments requires a means for non-intrusive and automated monitoring in ubiquitous en-vironment.

Common symptom caused by alcohol-effect on sensory-motor function can be observed in disturbed gait and balance, due to the influence in disrupting the transmission and integration of sensory and motor information [15]. For instance, Modig et al. reported that a 0.06 to 0.10 % BAC (Blood-Alcohol Content) resulted in decreased stability and sensorimotor adap-tation, as well as weakening the contribu-tion of visual information to postural con-trol [16]. Balance control is a complex pro-cess involving the integration of sensory input from multiple sensors, including vis-ual, vestibular, proprioceptors and mech-anoreceptors. The central nervous system processes these multiple sensory signals to estimate the position and motion of the body [17]. Therefore, an alcohol monitor-ing system could be able to monitor changes in gait as a function of sensory information available. In practice, visual observation of an individual’s ability to walk along a straight line is used as a test of soberness. A number of researchers have attempted to use sensors to objectively measure this field sobriety test. Klebe et al. measured and analyzed the influence of al-cohol consumption on gait in patients with essential tremor using a sensor-equipped treadmill in a laboratory setting [18].

Apart from alcohol monitoring, wear-able technology based on miniaturized sen-sors (e.g., accelerometer and gyroscope) is increasingly being used to monitor ac tivity and to find possible correlation to health-related conditions. Xu et al. have used ac-celerometers to monitor gait quality of

stroke survivors [19]. Adelsberger et al. per-formed automated Romberg testing for pa-tients with benign paroxysmal positional vertigo [20]. Gietzelt et al. evaluated gait pa-rameters measured by a single waist-mounted accelerometer during everyday life of patients with dementia [21], and Marschollek et al. developed unobtrusive method to determine individual fall risk based on the use of motion sensor data [22]. Palmerini et al. measured the acceler-ation of the low back to differentiate gait patterns in healthy adults and those with Parkinson’s Disease (PD) [23]. Salarian et al. proposed the use of accelerometers and gyroscopes to detect early signs of progres-sion of PD [24]. Bächlin et al. introduced the use of wearable assistant to help patients with symptoms of freezing of gait (FOG) re-lated to PD [25]. Mazilu et al. proposed a wearable system which supports training of patients with PD and FOG in daily living [26]. Gait for Parkinson’s disease appears definite on /off stages and a reduced smoothness and dynamics in trunk move-ment during off-stage gait [23], whereas, abnormal gait due to alcohol shows various pattern including discouraged walking, a wide-based gait and poor tandem gait [27].

Objective quantification of alcohol-in-duced gait was studied in recent researches. In [28], accelerometer in smartphone was utilized to detect change in alcohol-induced gait, and achieved 56 % accuracy in detec-tion. Another phone-based analysis system was developed to record the location and time context of alcohol-induced gait [29]. The inertial sensors in those systems are highly accessible through using smart-phones without any additional accessory to measure spatial-temporal param eters of gait. However, monitoring gait using inertial sen-sors has the problem of the high rate of error associated to these sensors. Acceleration is the derivative of velocity and involves high frequency components, and the orientation of gyroscope-based system suffers from drift over time [30 – 32]. Therefore, understand-ing of gait movements using inertial sensors requires intensive computational processing and external signals to re-calibrate the units from the drift [30].

To overcome noisy data from inertial sen-sors, more recent research and practical ap-plications have used pressure sensors to de-

tect gait events. Howell et al. introduced a mobile gait system using low-cost insole for stroke patients [33]. Bamber et al. used press-ure sensors to quantify movement of patients with Parkinson’s disease [34]. Electronic pressure monitoring systems, placed as an insole into regular shoes, have also been used to continuously monitor pres sure over the plantar surface to provide information on the process of ulcer formation [35].

In this study, we propose a remote moni-toring system leveraging machine learning and sensor technology. Our system analyzes alcohol-induced gait with person-dependent model based on the individual’s normal walking pattern that can be monitored and stored by sensor-equip ped shoes. Being combined to daily health monitoring, detec-tion of alcohol-induced gait can be built with wearable devices as an intelligent system that aims at empowering the capabilities of sys-tems or people by means of machine learn-ing and sensor technology [36, 37].

2. Objectives

The primary objective of this study was to build a personal gait-monitoring system, using sensorized smart shoe technology, for the detection of alcohol-induced changes in individuals’ gait pattern. More specifically, this work utilizes a pair of smart shoes equipped with an array of pressure sensors, which is capable of pro-viding in-depth understanding of spatio -temporal properties of human gait affected by alcohol. Our specific goals were to: (i) extract physiologically meaningful features of gait using biomechanical variables of the gait pattern, and (ii) demonstrate the accu-racy of our sensor system in detecting in-toxication-induced impairments in gait.

3. Methods3.1 Participants and Study ProtocolThe study was approved by UCLA institu-tional review board (IRB #13 – 000207). Twenty healthy participants (12 men and 8 women) volunteered for our study and pro-vided informed consent. The relevant char-acteristics of the study group were as fol-

For personal or educational use only. No other uses without permission. All rights reserved.Note: Uncorrected proof, epub ahead of print online

Downloaded from www.methods-online.com on 2016-10-28 | IP: 64.60.11.34

© Schattauer 2016 Methods Inf Med 6/2016

3

Wea

rabl

e Th

erap

y

E. Park et al.: Alcohol-impaired Gait Monitoring Using Smart Shoes

lows: ages of 24.78 ± 4.97, heights of 176.7 ± 11.40 cm, weights of 70.8 ± 14.65 kg, and foot lengths of 25.7 ± 1.9 cm. All partici-pants were above the legal drinking age, which is 21 in the United States. We instru-mented shoes of various sizes (M7, M8, M9.5, M11, and F6, F7, F8.5) to appropri-ately fit participants with their approximate shoe size for accurate localization of high-pressure regions under the foot.

Before the gait assessment, participants performed a sobriety-walking test to ensure the absence of alcohol-related effects on baseline measures. BAC was measured using a breath analyzer (BACtrack S80 Pro Breathalyzer Professional Edition). Any par-ticipant whose initial BAC was not 0.0 % was excluded from the study. Then, participants were assessed for their baseline performance by walking a 12m walkway at their preferred speed of normal walking. They were further asked to walk at slow and fast speed in order to obtain sober gait pattern that deviate from the baseline walking, i.e. sober walking at preferred speed. The target BAC indicat-ing alcohol intoxication was set to 0.06 % at which stability was affected as investigated in [16]. Participants consumed three to five glasses of 1.5 ounces of 40 % alcohol (Smir-noff Triple Distilled Vodka No. 21TM). After a resting period of 15 minutes to wash out alcohol remaining in the lining of the mouth, BAC levels were re-evaluated to en-sure the target BAC of 0.06 % had been reached. Once the BAC level was reached, participants indicative of alcohol intoxi-cation completed another set of gait trials to

measure the performance of alcohol-intoxi-cated gait.

3.2 Task Setup, Equipment and Monitoring Software

We developed the smart shoe shown in ▶ Figure 1, with a pressure sensor array embedded in the insole. An embedded processor (Arduino Fio) was used to collect and transmit the sensor data to a gateway computer in real-time via a Zigbee (IEEE 802.15) channel. The sample rate of data acquisition was 100 Hz. Pressure sensors

were placed in the insole for the detection of heel-pressure (H), mid-lateral plantar pressure (M), and big toe pressure (T). We developed a software program, Gait Per-formance Analysis System (GPAS), that receives and analyzes the transmitted gait data. GPAS was implemented in C #, .NET framework and MATLAB (▶ Figure 2).

3.3 Detection of Alcohol-induced Gait Disturbance

▶ Figure 3 shows a schematic representa -tion of the analysis process to quantify al-

Figure 1 The smart shoe sys-tem, containing an embedded processor and Zigbee receiver. Pressure sensors in the insole are used to de-tect the following gait pressure: heel-strike (H), mid-lateral plantar pressure (M), and big toe pressure (T).

Figure 2 Gait Performance Analysis System (GPAS): (a) GPAS connecting shoes and analysis software; (b) plantar-pressure map shows the pressure distribu-tion with contour plot of circles (at left heel contact) and sensor signal processing with pressure values used to detect signal peaks, segment gait events, and step-specific data.

a b

For personal or educational use only. No other uses without permission. All rights reserved.Note: Uncorrected proof, epub ahead of print online

Downloaded from www.methods-online.com on 2016-10-28 | IP: 64.60.11.34

Methods Inf Med 6/2016 © Schattauer 2016

4

Wea

rabl

e Th

erap

y

E. Park et al.: Alcohol-impaired Gait Monitoring Using Smart Shoes

walking with preferred speed and walking with various speeds. Then, the features of GAIT-ALCOHOL data were subtracted from the baseline features which were extracted from the individual’s normal walking at preferred speed, in order to re-flect individual’s various gait characteristics in daily life as baseline in model and vali-date the effect of such deviate in the de -tection. Similarly, the features of GAIT-SOBER were also subtracted from the base-line features. This process is referred to as the personalization process using personal gait profile (baseline). Then, a feature selec-tion algorithm followed by a classification (random forest) algorithm was employed to create a classification model to provide prediction on a testing data (GAIT-SOBER or GAIT-ALCOHOL). All components of the analysis process shown in ▶ Figure 3 are discussed in detail in the following sub-sections.

3.3.1 Gait Cycle Segmentation and Feature Extraction

The raw sensor data was used to construct a pressure distribution map to detect im-portant gait events, i.e. heel-strike and toe-off (▶ Figure 1). Notations related to im-portant gait events used in our analysis are summarized in ▶ Table 1.

A threshold-based peak detection algo-rithm was used to locate peak pressures at the heel and toe regions of each foot (see ▶ Figure 4). We applied noise-tolerant peak detection using threshold and selec-tive index of each peak, which determine local peaks if each peak is significantly larger than the data around it [38 – 40]. Peaks must be larger than a threshold value to be maxima, and selective index measure the amount above surrounding data for a peak to be identified. The peak detection algorithm tolerates to false peaks which would be filtered by threshold and selective index that are adaptively set to sensor sig-nal. Successive heel contacts, which gener-ated peak pressure signals from the heel sensors, were used to determine gait cycle time GCTi,j , defined as the temporal in -terval between two consecutive heel con-tacts of (i − 1)th and i th step and quantified as CTi,j (heel) − CTi − 1,j (heel) . Stance time, STNCi,j , was calculated as the time du-

Figure 3 Flow of the analysis process for the detection of alcohol-induced gait

Figure 4 Detection of gait events using pressure sensors on heels

Gait events

i

j

k

CTi,j (k)

CPi,j (k)

GCTi,j

STNCi,j

Description

i th step

Foot j {left, right}

Pressure sensor k {heel, mid-lateral, toe} for the con-tacts of plantar

Contact time of foot j for the i th step using sensor k

Pressure value of sensor k on foot j for the i th step

i th gait cycle time (GCT, stride time) of foot j

i th stance time of foot j

Table 1 Notations for the de-tection of gait events

cohol-induced gait impairments. The data obtained from alcohol-induced walking are labeled as GAIT-ALCOHOL (class 1) and the data obtained from sober walking with

speed variation labeled as GAIT-SOBER (class 0). First, the same set of gait features were extracted from all walking tests in-cluding alcohol-induced walking, normal

For personal or educational use only. No other uses without permission. All rights reserved.Note: Uncorrected proof, epub ahead of print online

Downloaded from www.methods-online.com on 2016-10-28 | IP: 64.60.11.34

© Schattauer 2016 Methods Inf Med 6/2016

5

(LOOCV), since the composition of train-ing set and test set sensitively determines the prediction of classifier [49 – 51]. The nested LOOCV technique contains two layers of LOOCV. First, the technique di-vides the entire set of data into a learning dataset and a validation dataset based on LOOCV. Then, the feature in the validation set is recursively eliminated, which per-forms additional LOOCV to compute the best feature set; the learning set is further divided into training and testing sets based on the inner LOOCV.

In summary, the analyses returned (i) the discrimination of alcohol-induced gait patterns from normal gait, and (ii) the ex-pected classification when the reported fea-ture sets are used.

4. Results4.1 Alcohol-induced Gait Disturbances in Individual SubjectsGAIT-ALCOHOL and GAIT-SOBER were compared for the representative features including SPEED, CADENCE, STRL, GCT-AVG, and SI (▶ Figure 5). Although

input, and the class with the most votes will be output as the predicted class from the RF.

RF intrinsically contains the feature se-lection mechanism as it randomly selects different features to construct each tree within its forest. However, there exist other works that have shown that RF combined with a feature selection algorithm can im-prove the results [45]. The purpose of fea-ture selection was to identify a set of gait fea-tures containing information that would discriminate walking behavior and reduce the searching space. We have conducted Wrapper Subset Evaluation to assess subsets of gait parameters according to their con-tributions to the classification performance [45, 46]. The association of features was suc-cessively evaluated and then the optimal set of features was iteratively discovered through this evaluation. The num ber of fea-tures is also known to affect the inter-tree correlation and the strength of the individu-al decision tree in random forest (RF) [47]. In this study, we employed a WEKA imple-mentation of the RF meth od [48].

In order to obtain more robust estimates of the classifier performance, we imple-mented a leave-one-out cross validation

Wea

rabl

e Th

erap

y

E. Park et al.: Alcohol-impaired Gait Monitoring Using Smart Shoes

ration between heel contact and the corre-sponding toe contact: CTi,j (toe) − CTi,j (heel) .

A total of 44 gait features of a walking phase, which is composed of consecutive gaits, were extracted from the sensor data, based on biomechanical properties of hu -man locomotion (▶ Table 2). In our ex-perimental setting, a walking phase is a 12-meter walk test, and features for current walking segment could be reflected to a personal profile which aggregates individ -ual’s normal-walking data.

3.3.2 Adjustment Using Personal Gait Profile

Extracted gait parameters were individual -ized by a two-step procedure. First, some of the parameters whose values are in-herently different due to individual anthro-pometrics were calibrated. These individu-ally calibrated parameters included GCT, STNC, and SSR, which were normalized to shoe size. Each gait feature was then ‘per-sonalized’ by being subtracted from the corresponding feature in baseline. Then, differences between baseline and alcohol-induced gait features were evaluated using paired-t test analyses. The adjusted gait fea-tures were used to build PDM (person- dependent model) to evaluate effects of al-cohol intoxication. We also compared gait performance based on PIM (person-inde-pendent model) that eliminates this adjust-ment to demonstrate the effect of personal-ization in Section 4.

3.3.3 Feature Selection and Random Forest Classification

To classify alcohol-induced gait features from normal variations in gait, we adopted a supervised machine-learning classification method. Among various machine learning algorithms commonly used in biomech-anics, we selected the RF classifier devel-oped by Leo Breiman [42, 43]. RF is an en-semble predictor that consists of multiple decision trees. The RF method, including variations on the RF classification method, has successfully been used in the fields of computer vision and medical informatics [43, 44]. To classify a new queried sample with an input vector, each decision tree in the forest predicts the output class from the

Table 2 Description of representative features for a walking segment in gait analysis

Feature

SPEED

STEPSj

CADENCE

STRL

GCT-AVG(j)GCT-VAR(j)GCT-SD(j)

STNC-AVG(j)STNC-VAR(j)STNC-SD(j)

SSR-AVG(j)SSR-VAR(j)SSR-SD(j)

SI

PV-AVG(j,k)PV-VAR(j,k)PV-SD(j,k)

SumPV(j)

Description

The average speed of a walk phase

The number of steps of foot j in a walk phase

The average number of steps per minute

The average of stride length

Mean value, variance, and standard deviation of gait cycle time (GCT) of foot j

Mean value, variance, and standard deviation of stance time (STNC) of foot j

Mean value, variance, and standard deviation of stance to stride ratio (SSR) of foot j (STNCi,j / GCTi,j)

Symmetry Index [41],

SI = × 100

Mean value, variance and standard deviation of pressure sensor k ’s peaks of foot j

Sum of pressure sensors on the insole of foot j

Unit

(m /s)

(steps / min)

(m)

(s)

(s)

(%)

For personal or educational use only. No other uses without permission. All rights reserved.Note: Uncorrected proof, epub ahead of print online

Downloaded from www.methods-online.com on 2016-10-28 | IP: 64.60.11.34

Methods Inf Med 6/2016 © Schattauer 2016

6

Wea

rabl

e Th

erap

y

E. Park et al.: Alcohol-impaired Gait Monitoring Using Smart Shoes

alcohol intoxication produced an overall decrease in SPEED , of – 0.059 m / s (SD 0.16 m / s), there was significant inter-individual variability with 13 (out of 20) participants (65 %) showing decrease in gait speed post-intoxication, whereas the remaining 35 % increased their gait speed with alcohol in-toxication. Individualized effects of alcohol

were also identified for other gait features, as shown in ▶ Figure 5.

4.2 Performance of Random Forest Classifier

The classification algorithm utilized spa-tiotemporal gait features, pressure-related

features, and personal gait profiles, nor-malized to foot size and baseline gait pat-tern. RF classifier performed with 86.2 % accuracy. ▶ Table 3 reports the classifica -tion performance with selected attributes.

We evaluated the performance of our classifier for different sets of inputs, namely by PDM for individualized gait model, PIM

Figure 5 Gait changes before and after alcohol intoxication

Person-Dependent Model

PIM

PIM

PIM

PIM

PIM

PIM

PDM

PDM

PDM

PDM

PDM

PDM

PDM = Person-Dependent Model; PIM = Person-Independent Model; STS = Spatial-Temporal Sensing; STPRS = STS plus Plantar Pressure Sensing; AG = Al-cohol-induced Gait; NG = Normal Gait

SensingModality

STS

STS

STS

STPRS

STPRS

STPRS

STS

STS

STS

STPRS

STPRS

STPRS

Target Class

AG

NG

Average

AG

NG

Average

AG

NG

Average

AG

NG

Average

Evaluation

Accuracy

65.0 %

77.5 %

79.3 %

86.2 %

Precision

0.500

0.731

0.650

0.778

0.774

0.775

0.789

0.800

0.794

0.842

0.900

0.866

Recall

0.500

0.731

0.650

0.500.

0.923

0.775

0.882

0.667

0.793

0.941

0.750

0.862

F-Measure

0.500

0.731

0.650

0.609

0.842

0.760

0.833

0.727

0.789

0.889

0.818

0.860

MCC

0.231

0.483

0.569

0.716

AUC

0.624

0.666

0.831

0.882

PRC Area

0.588

0.725

0.677

0.715

0.737

0.730

0.803

0.832

0.851

0.916

0.871

0.897

Table 3 Performance of classifier for alcohol-induced gait disturbances

For personal or educational use only. No other uses without permission. All rights reserved.Note: Uncorrected proof, epub ahead of print online

Downloaded from www.methods-online.com on 2016-10-28 | IP: 64.60.11.34

© Schattauer 2016 Methods Inf Med 6/2016

7E. Park et al.: Alcohol-impaired Gait Monitoring Using Smart Shoes

for monitoring without personal sensing profile, STS (SpatioTemporal Sensing) for sensing without pressure information and STPRS (STS plus Plantar Pressure Sensing) indicating sensor diversity including pres -sure. Performance with these different sets of input data is reported in ▶ Table 3. First, we excluded personal profile information and pressure information from the input set. Without both personal profiling and pressure values, the classifier exhibited an accuracy of 65 %. In a second round of test-ing of our classifier, classification with STPRS exhibited an accuracy of 77.5 %. The third round of evaluation with PDM and STS achieved an accuracy of 79.3 % in de-tection, and the classification with PDM and STPRS enhances the detection per-formance to an accuracy of 86.2 %.

5. Discussion

The proposed system aims to verify the de-tection capability of sensor-equipped shoes that is unobtrusive to use in daily life. The monitoring tool is feasible to be used in ubiquitous environment, since it is com -posed of normal shoes with sensors and wireless transmission to mobile gateway. Most wearable healthcare devices such as wristband, necklace, and shirts are de-signed to transmit the sensing data to mo-bile gateway including smart phones or home gateway.

Before performing analysis, we expected clear detection of alcohol-intoxicated gait with simple classifier. However, individual pattern of gait change under the influence was not uniform as shown in ▶ Figure 5. The importance of personal profile of ac-tivity in daily living is increased, and the re-sult corresponds to the trend of healthcare systems in which data is more important than algorithm.

5.1 Alcohol Intoxication and Gait Performance

To circumvent the limitations of inertial sensors in this application, we have de-signed a smart shoe system, which uses pressure sensor data to identify changes in the spatial-temporal and pressure profiles of gait induced by alcohol intoxication (see

▶ Table 3). Using a stepwise analysis of the performance of our RF classifier, we dem-onstrated the added contribution of using pressure information in combination with spatial-temporal information, which im-proved the precision and recall of classifi-cation by 19.2 % (from 0.650 to 0.775) with PIM. With PDM, combination of pressure values with spatial-temporal features achieves 9.1 % higher precision (from 0.794 to 0.866) and 8.7 % higher recall (from 0.793 to 0.862). In addition, the diversity of sensors affects more to PIM than PDM in all evaluation metrics.

This improvement in classifier perform-ance results from the specific information on initial and terminal contact that can be extracted from pressure data independent of the pattern and speed of gait [32, 52, 53]. The medical application of pressure-based gait monitoring has been also demon-strated in [17, 36, 47].

5.2 Detection with Characteristics of Personal Gait Pattern

In addition to the use of multiple features in machine learning to detect alcohol- induced gait, we measured the effect of personalization of gait performance. It is generally known that alcohol prolongs the latency and reduces the amplitude of long latency muscle responses due to its dam-pening effect over mono- and poly-synap-tic reflexes [54, 55]. However, other studies observed the ‘individualized’ and ‘variable’ nature of gait behaviors [36, 56 – 58], de-monstrating the importance of providing the classifier system with prior information on an individual’s usual walking behavior during the training phase.

Our experimental outcomes also dem-onstrated that the effects of alcohol on gait speed and cadence were individualized (▶ Figure 5). The a priori training allows the classifier to adjust the detection of alco-hol-induced changes in gait to personal (baseline) gait behavior. Personalization of the classifier significantly improved detec-tion of the intoxicated state, increasing the accuracy of the classifier by 11.2 % (from 0.775 to 0.862), when pressure value analy-sis was applied. Without pressure value analysis, accuracy was enhanced by 22 % (from 0.650 to 0.793). We believe that the

main reason for the poor classifier per-formance without personalization is the high degree of within-subject variation in gait performance under different con-straints [24, 31, 59]. We must also consider individual differences in alcohol tolerance, such that gait may have been minimally af-fected by intoxication levels of 0.06. These findings have motivated us to further de-velop our smart shoe system to monitor everyday walking in an attempt to capture the variation in gait characteristics, both within and between individuals.

5.3 Objectively-monitored Gait for Quantified Self

Our results provide valuable insight re-garding the requirement for monitoring technology to take into account affecting factors including type of sensors and vari-ous factors belonged to participants in the analysis [36]. As an example, we believe that the detection rate of intoxication could be significantly improved by combining measured gait patterns with information on the location, sensory environment, or an individual’s past alcohol history (e.g., chronic alcoholism or alcohol related of-fence). Furthermore, demographic infor-mation, such as sex, could further improve the detection of alcohol intoxication, as women are known to be less tolerant to al-cohol compared to men. In the future, a larger study can be performed to further investigate the contribution of personal demographic information, such as the length of lower limbs [24], sex [31], and associated diseases [59]. We believe that the adequate personalizing the model will improve the detection rate.

The most significant improvement pro-vided by our system is the objective, real-time monitoring of gait using an incon-spicuous system. Patients who require complete abstinence would be able to avoid the inconvenience and discomfort of re-peated BAC measurement at regular inter-vals. In addition, gait data would be more robust for the identification of intoxication than self-report or the horizontal gaze nyst-agmus test, which would be fraudulent [60]. We also have advanced knowledge and application of wearable devices, pro-posing an alternative to common devices

Wea

rabl

e Th

erap

y

For personal or educational use only. No other uses without permission. All rights reserved.Note: Uncorrected proof, epub ahead of print online

Downloaded from www.methods-online.com on 2016-10-28 | IP: 64.60.11.34

Methods Inf Med 6/2016 © Schattauer 2016

8 E. Park et al.: Alcohol-impaired Gait Monitoring Using Smart Shoes

that are heavy and highly obtrusive [61]. Our smart shoe device performs similar to patch or bandage type sensors [62, 63] or watch-shaped activity monitors [64, 65] that are easy-to-use devices.

In this study, we have investigated the detection of intoxication solely based on walking behaviors using a wearable system, and showed that the results are very prom-ising. In addition, the system is unobtrus-ive, which is a necessary requirement for monitoring alcohol-related patients or criminals without constraints of place and time. The longitudinal collection of indi-vidual gait pattern using smart shoes may enable personalized BAC estimation model. We believe that this study opens new research opportunities for the remote monitoring of alcohol-related activities, which have not yet been actively investi-gated.

6. Conclusion

We have introduced a machine-learning based method for monitoring alcohol- induced impairments in gait using sen -sor-equipped smart shoes. This study in-vestigated and demonstrated the feasibility of detection of intoxication using gait monitoring. We have found that personal-ized monitoring and a machine-learning technique significantly improved the detec-tion rate of alcohol-injected gait activities. We believe that this study enables further research opportunities in the field of alco-hol-related activities monitoring which requires easy-to-use and intelligent deci-sion. Alcohol sensing with wearable de-vices enables participants’ adherence to monitoring, and can be utilized in various applications including customized treat-ment for patients.

Author Contribution

Conceived and designed the experiments: EP, SIL, DL, MS. Performed the experi-ments: EP, SIL, JG, AH, MA, AC, NG. Ana-lyzed the data: EP, SIL, HN. Contributed reagents / materials / analysis tools: EP, SIL, HN, SP, HC. Wrote the paper: EP, SIL, HN. Revised manuscript for important intellec-tual content: HC, DL, MS.

References1. Rehm J, Mathers C, Popova S, Thavorncharoensap

M, Teerawattananon Y, Patra J. Global burden of disease and injury and economic cost attributable to alcohol use and alcohol-use disorders. The Lan-cet. 2009; 373(9682): 2223–33.

2. Stahre M. Contribution of excessive alcohol con-sumption to deaths and years of potential life lost in the United States. Prev Chronic Dis. 2014; 11: E109.

3. NIH National Institute on Alcohol Abuse and Al-coholism (NIAAA), Alcohol Facts and Statistics 2016 [updated 2016 Jan; cited 2016 Apr 17]. Avail-able from: http://www.niaaa.nih.gov/alcohol-health/overview-alcohol-consumption/alcohol-facts-and-statistics.

4. Sokol RJ, Clarren SK. Guidelines for use of termi-nology describing the impact of prenatal alcohol on the offspring. Alcoholism: Clinical and Experi-mental Research. 1989; 13(4): 597–8.

5. Singal A, Jampana SC. Current management of al-coholic liver disease. Current Drug Abuse Re-views. 2013; 7(1): 18–28.

6. Wu L-T, Ringwalt CL. Alcohol dependence and use of treatment services among women in the community. Am J Psychiatry. 2004; 161(10): 1790–7.

7. Zeisser C, Stockwell TR, Chikritzhs T, Cherpitel C, Ye Y, Gardner C. A Systematic Review and Meta Analysis of Alcohol Consumption and Injury Risk as a Function of Study Design and Recall Period. Alcoholism: Clinical and Experimental Research. 2013; 37(s1): E1-E8.

8. Martin SE. The epidemiology of alcohol-related interpersonal violence. Alcohol Research and Health. 1992; 16(3): 230.

9. Conner KR, Huguet N, Caetano R, Giesbrecht N, McFarland BH, Nolte KB, et al. Acute use of alco-hol and methods of suicide in a US national sample. Am J Public Health. 2014; 104(1): 171–8.

10. Taylor B, Irving H, Kanteres F, Room R, Borges G, Cherpitel C, et al. The more you drink, the harder you fall: a systematic review and meta-analysis of how acute alcohol consumption and injury or col-lision risk increase together. Drug and Alcohol De-pendence. 2010; 110(1): 108–16.

11. Johnson NB, Hayes LD, Brown K, Hoo EC, Ethier KA. CDC National Health Report: leading causes of morbidity and mortality and associated beha-vioral risk and protective factors – United States, 2005–2013. MMWR Surveill Summ. 2014; 63(Suppl 4): 3–27.

12. Barnett NP. Alcohol sensors and their potential for improving clinical care. Addiction (Abingdon, England). 2015; 110(1): 1.

13. Marques PR, McKnight AS. Field and laboratory alcohol detection with 2 types of transdermal de-vices. Alcohol Clin Exp Res. 2009; 33(4): 703–11.

14. Lukowicz P, Kirstein T, Tröster G. Wearable Sys-tems for Health Care Applications. Methods Archive. 2004; 43(3): 232–8.

15. Fein G, Smith S, Greenstein D. Gait and Balance in Treatment Naïve Active Alcoholics with and with-out a Lifetime Drug Codependence. Alcoholism: Clinical and Experimental Research. 2012; 36(9): 1550–62.

16. Modig F, Fransson P-A, Magnusson M, Patel M. Blood alcohol concentration at 0.06 and 0.10 % causes a complex multifaceted deterioration of body movement control. Alcohol. 2012; 46(1): 75–88.

17. Modig F, Patel M, Magnusson M, Fransson P-A. Study II: Mechanoreceptive sensation is of in-creased importance for human postural control under alcohol intoxication. Gait & Posture. 2012; 35(3): 419–27.

18. Deuschl G. Influence of alcohol on gait in patients with essential tremor. Neurology. 2005; 65: 96101.

19. Xu X, Batalin M, Wang Y, Kaiser W, editors. Gait quality evaluation method for post-stroke patients. Acoustics, Speech and Signal Processing (ICASSP), 2012 IEEE International Conference on; 2012: IEEE.

20. Adelsberger R, Valko Y, Straumann D, Tröster G. Automated Romberg Testing in Patients with Be-nign Paroxysmal Positional Vertigo and Healthy Subjects. IEEE Trans Biomed Eng. 2015; 62(1): 373–81.

21. Gietzelt M, Wolf KH, Kohlmann M, Marschollek M, Haux R. Measurement of Accelerometry-based Gait Parameters in People with and without De-mentia in the Field. A Technical Feasibility Study. Meth Inf Med. 2013; 52(4): 319–25.

22. Marschollek M, Rehwald A, Wolf KH, Gietzelt M, Nemitz G, Meyer zu Schwabedissen H, et al. Sen-sor-based Fall Risk Assessment – an Expert ‘to go’ . Meth Inf Med. 2011; 50(5): 420–6.

23. Palmerini L, Mellone S, Avanzolini G, Valzania F, Chiari L. Quantification of motor impairment in Parkinson’s disease using an instrumented timed up and go test. IEEE Trans Neural Syst Rehabil Eng. 2013; 21(4): 664–73.

24. Salarian A, Zampieri C, Horak FB, Carlson-Kuhta P, Nutt JG, Aminian K, editors. Analyzing 180 turns using an inertial system reveals early signs of progression of Parkinson’s disease. Engineering in Medicine and Biology Society, 2009 EMBC 2009 Annual International Conference of the IEEE; 2009: IEEE.

25. Bächlin M, Plotnik M, Roggen D, Giladi N, Haus-dorff JM, Tröster G. A Wearable System to Assist Walking of Parkinson’s Disease Patients. Benefits and Challenges of Context-triggered Acoustic Cueing. Meth Inf Med. 2010; 49(1): 88–95.

26. Mazilu S, Blanke U, Hardegger M, Tröster G, Gazit E, Hausdorff JM, editors. GaitAssist: a daily-life support and training system for parkinson’s dis-ease patients with freezing of gait. Proceedings of the 32nd annual ACM conference on Human fac-tors in computing systems; 2014: ACM.

27. Sanders RD, Gillig PM. Gait and its assessment in psychiatry. Psychiatry (Edgemont). 2010; 7(7): 38–43.

28. Arnold Z, Larose D, Agu E, editors. Smartphone Inference of Alcohol Consumption Levels from Gait. Healthcare Informatics (ICHI), 2015 Inter-national Conference on; 2015: IEEE.

29. Kao H-LC, Ho B-J, Lin AC, Chu H-H, editors. Phone-based gait analysis to detect alcohol usage. Proceedings of the 2012 ACM Conference on Ubi-quitous Computing; 2012: ACM.

30. Woodman OJ. An introduction to inertial navi-gation. University of Cambridge, Computer Lab-oratory, Tech Rep UCAMCL-TR-696. 2007; 14: 15.

Wea

rabl

e Th

erap

y

For personal or educational use only. No other uses without permission. All rights reserved.Note: Uncorrected proof, epub ahead of print online

Downloaded from www.methods-online.com on 2016-10-28 | IP: 64.60.11.34

© Schattauer 2016 Methods Inf Med 6/2016

9E. Park et al.: Alcohol-impaired Gait Monitoring Using Smart Shoes

31. Lopez-Meyer P, Fulk GD, Sazonov ES. Automatic detection of temporal gait parameters in post-stroke individuals. IEEE Trans Inf Technol Biomed. 2011; 15(4): 594–601.

32. Fischer C, Talkad Sukumar P, Hazas M. Tutorial: implementation of a pedestrian tracker using foot-mounted inertial sensors. IEEE Pervasive Com-puting. 2013; 12(2): 17–27.

33. Howell AM, Kobayashi T, Hayes HA, Foreman KB, Bamberg SJ. Kinetic gait analysis using a low-cost insole. IEEE Trans Biomed Eng. 2013; 60(12): 3284–90.

34. Bamberg SJM, Benbasat AY, Scarborough DM, Krebs DE, Paradiso JA. Gait analysis using a shoe-integrated wireless sensor system. IEEE Trans Inf Technol Biomed. 2008; 12(4): 413–23.

35. Helbostad JL, Leirfall S, Moe-Nilssen R, Sletvold O. Physical fatigue affects gait characteristics in older persons. J Gerontol A Biol Sci Med Sci. 2007; 62(9): 1010–5.

36. Acampora G, Cook DJ, Rashidi P, Vasilakos AV. A survey on ambient intelligence in healthcare. Proc IEEE Inst Electr Electron Eng. 2013; 101(12): 2470–94.

37. Cook DJ, Augusto JC, Jakkula VR. Ambient intelli-gence: Technologies, applications, and opportun-ities. Pervasive and Mobile Computing. 2009; 5(4): 277–98.

38. Yoder N. PeakFinder. Available from: http://www.mathworks.com/matlabcentral/fileex-change/25500. 2011.

39. Khan T, Westin J, Dougherty M. Motion cue analysis for parkinsonian gait recognition. Open Biomed Eng J. 2013; 7: 1–8.

40. Adams D. Health monitoring of structural materi-als and components: Methods with applications. New York: John Wiley & Sons; 2007.

41. Anna AS, Wickström N. A symbol-based ap-proach to gait analysis from acceleration signals: Identification and detection of gait events and a new measure of gait symmetry. IEEE Trans Inf Technol Biomed. 2010; 14(5): 1180–7.

42. Breiman L. Random forests. Machine Learning. 2001; 45(1): 5–32.

43. Zhang N, Li B-Q, Gao S, Ruan J-S, Cai Y-D. Com-putational prediction and analysis of protein γ-carboxylation sites based on a random forest method. Molecular BioSystems. 2012; 8(11): 2946–55.

44. Li B-Q, Hu L-L, Chen L, Feng K-Y, Cai Y-D, Chou K-C. Prediction of protein domain with mRMR feature selection and analysis. PLoS One. 2012; 7(6): e39308.

45. Janecek A, Gansterer WN, Demel M, Ecker G, edi-tors. On the Relationship Between Feature Selec-tion and Classification Accuracy. FSDM; 2008: Citeseer.

46. Hall MA, Smith LA. Practical feature subset selec-tion for machine learning. In: McDonald C, editor. Computer Science ’98 Proceedings of the 21st Aus-tralasian Computer Science Conference ACSC’98, Perth, 4–6 February 1998. Berlin: Springer; 1998. p. 181–191.

47. Strobl C, Boulesteix A-L, Kneib T, Augustin T, Zei-leis A. Conditional variable importance for ran-dom forests. BMC bioinform. 2008; 9(1): 1.

48. Hall M, Frank E, Holmes G, Pfahringer B, Reute -mann P, Witten IH. The WEKA data mining soft-ware: an update. ACM SIGKDD Explorations Newsletter. 2009; 11(1): 10–8.

49. Cawley GC, Talbot NL. Efficient leave-one-out cross-validation of kernel fisher discriminant classifiers. Pattern Recognition. 2003; 36(11): 2585–92.

50. Chuang L, Yang C, Wu K, Yang C. Correlation-based gene selection and classification using Tagu-chi-BPSO. Meth Inf Med. 2010; 49(3): 254.

51. Lee SI, Ghasemzadeh H, Mortazavi BJ, Sarrafza-deh M. A pervasive assessment of motor function: A lightweight grip strength tracking system. IEEE J Biomed Health Inform. 2013; 17(6): 1023–30.

52. Ostadabbas S, Nourani M, Saeed A, Yousefi R, Pompeo M. A Knowledge-Based Modeling for Plantar Pressure Image Reconstruction. IEEE Trans Biomed Eng. 2014; 61(10): 2538–49.

53. Giacomozzi C, Uccioli L. Learning from experi-ence: A simple effective protocol to test footwear prescriptions for the Diabetic foot by using the Pedar system. Journal of Biomedical Science and Engineering. 2013; 6: 45–57.

54. Ashby P, Carlen PL, MacLeod S, Sinha P. Ethanol and spinal presynaptic inhibition in man. Annals of Neurology. 1977; 1(5): 478–80.

55. Ahmad S, Rohrbaugh JW, Anokhin AP, Sirevaag EJ, Goebel JA. Effects of lifetime ethanol consump-tion on postural control: a computerized dynamic posturography study. Journal of Vestibular Re-search. 2002; 12(1): 53–64.

56. Nessler JA, Gilliland SJ. Interpersonal synchroni -zation during side by side treadmill walking is in-fluenced by leg length differential and altered sen-sory feedback. Human Movement Science. 2009; 28(6): 772–85.

57. Chen G, Patten C, Kothari DH, Zajac FE. Gait dif-ferences between individuals with post-stroke hemiparesis and non-disabled controls at matched speeds. Gait & Posture. 2005; 22(1): 51–6.

58. Yamasaki M, Sasaki T, Torii M. Sex difference in the pattern of lower limb movement during tread-mill walking. Eur J Appl Physiol Occup Physiol. 1991; 62(2): 99–103.

59. Arkadir D, Louis ED. The gait and balance dis-order of essential tremor: what does this mean for patients? Ther Adv Neurol Disord. 2013; 6(4): 229–236.

60. Booker J. The Horizontal Gaze Nystagmus test: Fraudulent science in the American courts. Science & Justice. 2004; 44(3): 133–9.

61. Zheng Y-L, Ding X-R, Poon CCY, Lo BPL, Zhang H, Zhou X-L, et al. Unobtrusive sensing and wear-able devices for health informatics. IEEE Trans Biomed Eng. 2014; 61(5): 1538–54.

62. Selvaraj N, Narasimhan R, editors. Automated prediction of the apnea-hypopnea index using a wireless patch sensor. Engineering in Medicine and Biology Society (EMBC), 2014 36th Annual International Conference of the IEEE; 2014: IEEE.

63. Shin K, Park G, Kim JP, Lee T, Ko BH, Kim YH, editors. An ultra-low power (ULP) bandage-type ECG sensor for efficient cardiac disease manage-ment. Conference proceedings: Annual Inter-national Conference of the IEEE Engineering in Medicine and Biology Society; 2012.

64. Takacs J, Pollock CL, Guenther JR, Bahar M, Na-pier C, Hunt MA. Validation of the Fitbit One ac-tivity monitor device during treadmill walking. Journal of Science and Medicine in Sport. 2014; 17(5): 496–500.

65. Bai Y, Welk GJ, Nam YH, Lee JA, Lee J-M, Kim Y, et al. Comparison of Consumer and Research Monitors under Semistructured Settings. Medi-cine and Science in Sports and Exercise. 2016; 48(1): 151–8.

Wea

rabl

e Th

erap

y

For personal or educational use only. No other uses without permission. All rights reserved.Note: Uncorrected proof, epub ahead of print online

Downloaded from www.methods-online.com on 2016-10-28 | IP: 64.60.11.34