1 of 4

Abstract—The objective of this pilot work is to identify characteristics and measure severity of motor fluctuations in patients with Parkinson’s disease (PD) based on wearable sensor data. Improved methods of assessing longitudinal changes in PD would enable optimization of treatment and maximization of patient function. We hypothesize that motor fluctuations accompanying late-stage PD present with predictable features of accelerometer signals recorded during execution of standardized motor tasks. Six patients (age 46-75) with diagnosis of idiopathic PD and levodopa-related motor fluctuations were studied. Subjects performed motor tasks in a “practically-defined OFF” state, and then at 30 minute intervals after medication intake. At each interval, data from 8 uniaxial accelerometers on the upper and lower limbs were recorded continuously, and subjects were videotaped. Features representing motion characteristics such as intensity, rate, regularity, and coordination were derived from the sensor data, and clinical scores were assigned for each task by review of the videotapes. Cluster analysis was performed on feature sets that were expected to reflect severity of parkinsonian symptoms (e.g. bradykinesia) and motor complications (e.g. dyskinesias). Two-dimensional data projections revealed clusters corresponding to the degree of dyskinesia and bradykinesia indicated by clinical scores. These preliminary results support our hypothesis that wearable sensors are sensitive to changing patterns of movement throughout the medication intake cycle, and that automated recognition of motor states using these recordings is feasible.

Keywords—Parkinson’s disease, wearable technology,

Sammon’s map

I. INTRODUCTION

Parkinson’s Disease (PD) is the most common disorder of movement, affecting about 3% of the population over the age of 65 years and more than 500,000 US residents. Its characteristic motor features are development of rest tremor, bradykinesia, rigidity, and impairment of postural balance. The primary biochemical abnormality in PD is deficiency of dopamine due to degeneration of neurons in the substantia nigra pars compacta. Current therapy of PD is based primarily on augmentation or replacement of dopamine, using the biosynthetic precursor levodopa or other drugs which activate dopamine receptors [13]. These therapies are often successful for some time in alleviating the abnormal movements, but most patients eventually develop motor complications as a result [3][9]. The complications include wearing off, the abrupt loss of efficacy at the end of each

dosing interval, and dyskinesias, involuntary and sometimes violent writhing movements. Wearing off and dyskinesias produce substantial disability, and frequently prevent effective therapy of the disease [6][7]. Although wearing off and dyskinesias often appear related to the timing of medication doses, they are not simply a consequence of the pharmacokinetics of levodopa. Motor complications are virtually never observed early in the treatment of PD; they appear only after prolonged treatment, usually several years. Experiments using controlled administration of dopaminergic drugs support these clinical observations [2][8].

The currently available tools for monitoring and managing motor fluctuations are quite limited. In clinical practice, information about motor fluctuations is usually obtained by asking the patient to recall the number of hours of ON and OFF time they have experienced in the recent past. This kind of self-report is subject to both perceptual bias (difficulty of distinguishing dyskinesia from other symptoms) as well as recall bias. Another approach is the use of patient diaries, which does improve reliability by recording symptoms as they occur, but does not capture many of the features that are useful in clinical decision-making [10]. In clinical trials of new therapies, the diary-based approach as well as extended direct observations of patients in a clinical care setting [1] have been used, but both capture only a small portion of the patients’ daily experience and are burdensome. A reliable quantitative tool for evaluating motor complications in PD patients would be valuable for routine clinical care of patients as well as for trials of novel therapies. In routine care, it would be very useful to obtain information on a patient’s motor pattern during the course of several days, and relate this to the timing and dose of medications. This would facilitate the planning of changes in treatment to minimize adverse symptoms. An even more important application would be for use in randomized trials of novel therapies for PD. By providing an accurate and unbiased measure of motor state, such a tool would likely reduce the number of subjects required and the duration of treatment required to observe an effect in a trial of a new agent. This in turn would translate into reduced cost and greater efficiency in the testing and development of therapies. Wearable sensor technology offers the tools to monitor motor functions in PD patients. Also, the interest in wearable systems is growing and is expected to lead to

Advanced Analysis of Wearable Sensor Data to Adjust Medication Intake in Patients with Parkinson’s Disease

Delsey M. Sherrill, MS 1, Richard Hughes, PT 1, Sara S. Salles, DO 5, Theresa Lie-Nemeth, MD 1, Metin Akay, PhD 3, David G. Standaert, MD, PhD 2, and Paolo Bonato, PhD 1,4

1 Dept of Physical Medicine & Rehabilitation, Harvard Medical School, Spaulding Rehabilitation Hospital, Boston, MA; 2 Dept of Neurology, Harvard Medical School, Massachusetts General Hospital, Boston, MA; 3 Thayer School of Engineering, Dartmouth College, Hanover, NH; 4 The Harvard-MIT Division of Health Sciences and Technology, Cambridge, MA; 5 Dept of Physical Medicine and Rehabilitation, University of Kentucky, Lexington, KY.

2 of 4

systems in the near future that can be worn continuously like a normal garment. However, before wearable systems can be used in clinical management of patients with PD, it is necessary to develop data analysis techniques that identify characteristics of motor patterns associated with the severity of symptoms and motor complications. This paper presents preliminary evidence that features can be derived from accelerometer data that correlate with clinical assessments, thereby enhancing the potential of wearable systems to enable adjustments in medication intake on the basis of quantitative measures.

II. METHODOLOGY A. Data Collection

Six individuals were recruited for the study, ranging in age from 46 to 75, with a diagnosis of idiopathic Parkinson’s disease (Hoehn & Yahr stage 2.5 to 3, i.e. mild to moderate bilateral disease with ability to recover from sudden postural disturbance or with some postural instability). Subjects were assessed on the day of the experiment by an experienced clinician using the Unified Parkinson’s Disease Rating Scale (UPDRS) [4]. Total scores over all subsections of the UPDRS ranged between 28-49, with a mean of 38 out of 199 points possible. Subjects were referred to the study based on a documented history of levodopa responsiveness, marked ON/OFF fluctuations, moderate to severe dyskinesia, bradykinesia, akinesia, and/or rigidity.

Miniature accelerometer sensors were used to gather biomechanical signals during standardized clinical tasks including sitting, finger-to-nose, tapping, sit-to-stand, walking, and stand-to-sit. For each task, 30 seconds of sensor data were recorded. Accelerometers were placed on the right and left upper arm, right and left forearm, right and left thigh, right shin, and sternum. The sensors were connected to an ambulatory system (Vitaport 3, Temec BV, the Netherlands) equipped with data acquisition hardware and software to collect and store the signals. Subjects were videotaped throughout the experiment, and motor UPDRS and dyskinesia scores were assigned for each task by review of the videotapes.

Subjects were asked to delay their first medication intake in the morning so that they could be tested in a “practically-defined OFF” state, the period of most severe parkinsonian symptoms. Subsequently, patients took their medications and were tested at 30-minute intervals thereafter, in order to gather sensor data during the “best ON” state, i.e. the period of maximal therapeutic benefit from medication, as well as during the “wearing off” period when dyskinesias tend to occur.

B. Feature Extraction

A set of ten 5-s epochs were randomly selected from the 30 s of sensor data corresponding to each task, for each pre- and post-dosage testing interval. To facilitate visualization

of the data, it was necessary to reduce the dimensionality of each epoch by selecting features that capture the characteristic accelerometer patterns associated with motor fluctuations. The features were chosen to represent characteristics such as intensity, modulation, rate, regularity, and coordination of movement. Intensity was measured as the root-mean-square (RMS) value of the detrended accelerometer signal. The modulation of the output of each sensor was used to represent dynamic characteristics of the tasks, and was calculated as the range of the autocovariance of each channel. Large values of this feature were indicative of intervals of rapid movements interspersed with intervals of slow movements. Rate of movement was represented by the dominant frequency component below 10 Hz. Regularity (periodicity) was measured by computing the ratio of energy in the dominant frequency component to the total energy below 10 Hz, the value of which approaches one for purely periodic signals. Coordination between body segments on the left and right side was captured in two aspects: magnitude (obtained by calculating the correlation coefficient) and delay (estimated as the time lag corresponding to the peak of the cross-correlation function).

To study movement patterns related to bradykinesia, features were derived from the upper extremity accelerometer channels during the task of “rapid alternating hand movements,” a task considered likely to evoke such patterns. To study dyskinetic movements, features were derived from the lower extremity accelerometer channels during a task requiring fine motor control of the upper extremity (“finger tapping”). This was done because the dyskinesias are more likely to be observed (i.e. less prone to suppression) in the lower extremities when the patient’s attention is focused on a task involving the upper extremities.

C. Data visualization

Further processing was necessary in order to visualize the data and assess whether data from the sensors was meaningfully related to clinical scores and/or the time course of the experiment. Suitable approaches to projecting high-dimensional data sets into a low-dimensional space include principal components analysis (PCA) [5] and Sammon’s mapping [11]. PCA involves computing the eigenvectors of the covariance matrix for the full data set and then sorting them according to their associated eigenvalues. Eigenvectors corresponding to the largest eigenvalues are the ones that account for the greatest amount of variation in the data set, i.e. they represent the principal components. The Sammon’s map algorithm reduces the dimensionality of the data set through an iterative procedure to optimize the preservation of euclidean distances between all pairs of data points. This projection technique was selected for its ability to preserve structure of the “data cloud” as it exists in the original feature space. In our analysis, both approaches were incorporated. In order to reduce computational complexity and minimize the

3 of 4

influence of redundant features, a PCA transformation was first applied to the normalized feature set, and the first 6 PCs (accounting for 90% of the total variance) were retained. Finally, Sammon’s map was applied to the PCA-transformed features to obtain a two-dimensional representation of the data.

III. RESULTS

The accelerometer time series shown in Figure 1

highlight the contrasting properties of ON and OFF states as captured in the sensor data. Relative improvements in amplitude, speed, and right-left coordination for movements of the upper limb during the ON state are apparent upon visual inspection of the signals. The observation of changes such as these guided the selection of features to associate with bradykinetic and dyskinetic movements.

Figures 2 and 3 show results of the two-dimensional projection of feature sets derived from sensor recordings as described previously. These data represent a complete motor fluctuation cycle of one individual. The variation of clinical scores over time is shown in the bar plot at the bottom of each figure. In the scatter plots, each point represents a single 5-second epoch, and its color indicates the clinical score assigned.

The spatial configuration of the points is a reflection of the similarity of the corresponding data epochs. Specifically, epochs with similar feature values will be co-located within a distinct region of the original feature space, and since the Sammon’s mapping preserves pairwise distances, those clusters will be apparent in the projection plane as well. Therefore, the color uniformity of a spatial cluster serves as an indicator of how strong a relationship exists between the clinical scores and wearable sensor data.

In Figure 2, the prominent cluster at right indicates that periods when dyskinesia is absent (score “0”) are clearly differentiated from periods when it is present. However the distinction between mild and moderate levels of severity is somewhat less clear-cut (scores of “1” and “2” respectively). A better understanding of each feature’s role in determining the structure of the data cloud would enable us to assess whether these distinctions are specific to dyskinetic movements or are a reflection of the amount of movement in general. This will be a principal aim of analyses in the near future.

In Figure 3, two main clusters are apparent. The cluster in the bottom right corner of the plane contains low scoring epochs (absent or mild bradykinesia) whereas the one at top left consists of epochs with high scores (mild, moderate, and severe bradykinesia). Further, the score-coloring of the clusters reveals an intriguing pattern: the spatial progression from lower right to upper left of the plane coincides with the ordering of scores in the UPDRS clinical scale. Such a pattern was especially clear in this case because the subject experienced more dramatic motor fluctuations than most others. It is an encouraging observation because it represents a first step toward validation of these measures.

-1

0

1

2

3

R a

rm

-1

0

1

2

3

L ar

m

-1

0

1

2

3

R le

g

0 1 2 3 4 5-1

0

1

2

3

L le

g

Time (s)

OFF medication

-1

0

1

2

3

R a

rm

-1

0

1

2

3

L ar

m

-1

0

1

2

3

R le

g

0 1 2 3 4 5-1

0

1

2

3

L le

g

Time (s)

ON medication

-1

0

1

2

3

R a

rm

-1

0

1

2

3

L ar

m

-1

0

1

2

3

R le

g

0 1 2 3 4 5-1

0

1

2

3

L le

g

Time (s)

-1

0

1

2

3

R a

rm

-1

0

1

2

3

L ar

m

-1

0

1

2

3

R le

g

0 1 2 3 4 5-1

0

1

2

3

L le

g

Time (s)

OFF medication

-1

0

1

2

3

R a

rm

-1

0

1

2

3

L ar

m

-1

0

1

2

3

R le

g

0 1 2 3 4 5-1

0

1

2

3

L le

g

Time (s)

-1

0

1

2

3

R a

rm

-1

0

1

2

3

L ar

m

-1

0

1

2

3

R le

g

0 1 2 3 4 5-1

0

1

2

3

L le

g

Time (s)

ON medication

Figure 1 Comparison of accelerometer data for ON and OFF medicationstates during the task of “alternating hand movements” in which the subjectalternately pronates and supinates the wrist as rapidly as possible and with maximal amplitude while in a seated position. Accelerations are shown inunits of gravity (g). Note differences in amplitude, speed, and coordinationof upper extremity movements as well as abnormal movements of thelower extremity.

012

0 9030 18015012060

DyskinesiaScore

Time (min)

LevodopaIntake

OFF

WEARING-OFFON

OFFMotor FluctuationCycle

Sammon’s map of Dyskinesia-related features

012

0 9030 18015012060

DyskinesiaScore

Time (min)

LevodopaIntake

OFF

WEARING-OFFON

OFFMotor FluctuationCycle

012

0 9030 18015012060

012

012

0 9030 180150120600 9030 18015012060

DyskinesiaScore

Time (min)

LevodopaIntake

OFF

WEARING-OFFON

OFF

LevodopaIntake

LevodopaIntake

OFF

WEARING-OFFON

OFFOFF

WEARING-OFFON

OFFMotor FluctuationCycle

Sammon’s map of Dyskinesia-related features

Figure 2 Two-dimensional projection of dyskinesia-related featuresfor a representative subject. Axis units are unspecified because theyresult from abstract transformations of normalized data and hence lacka direct physical interpretation. Scatter plot is color-coded accordingto clinical score assigned for each testing interval, as shown in thetimeline (0=absent, 1=mild, 2=moderate). A schematic of the motorfluctuation cycle is shown for reference.

4 of 4

IV. CONCLUSION

The results of this preliminary study indicate that

features estimated from accelerometer data can capture distinct patterns of movement associated with motor fluctuations in patients with PD. The study demonstrates a correlation between patterns identified via analysis of sensor data and clinical scores assigned by a physician experienced in the diagnosis and management of patients with PD. This finding is relevant to PD management, particularly in the adjustment of medication intake and the testing of new pharmacological treatments.

Clinicians encounter substantial difficulties in assessing motor fluctuations in patients with PD because of the limited reliability and objectivity with which patients perceive their condition. A wearable system equipped with accelerometers positioned on different body segments would address the need for gathering objective data to measure the response of symptoms and motor complications to changes in medication intake or the use of a new pharmacological agent.

In future analysis of motor fluctuations, we intend to study the interaction among symptoms and motor

complications (e.g., the simultaneous occurrence of rigidity and bradykinesia). Specifically, we plan to jointly track multiple symptoms and motor complications instead of considering them in isolation. Still further, we intend to expand the feature extraction procedures to include both linear and non-linear techniques. The latter have been demonstrated as having high sensitivity to motor abnormalities similar to those observed in patients with PD [12]. Finally, we will explore the use of neural networks and other pattern recognition algorithms to automatically identify fluctuations in severity of symptoms and motor complications. These advances will eventually lead to a wearable system capable of augmenting the ability of clinicians to optimally modify medication intake in patients with PD as the disease progresses.

ACKNOWLEDGMENT

This work was supported by the National Institute of

Neurological Disorders and Stroke, National Institutes of Health under the grant #R21NS045410-01A1.

REFERENCES

[1] Adler CH, Singer C, O'Brien C, Hauser RA, Lew MF, Marek KL, Dorflinger E, Pedder S, Deptula D, Yoo K, “Randomized, placebo-controlled study of tolcapone in patients with fluctuating Parkinson disease treated with levodopa-carbidopa. Tolcapone Fluctuator Study Group III”, Arch Neurol, 55(8): 1089-1095, 1998.

[2] Blanchet PJ, Papa SM, Metman LV, Mouradian MM, Chase TN, “Modulation of levodopa-induced motor response complications by NMDA antagonists in Parkinson's disease”, Neurosci Biobehav Rev, 21: 447-453, 1997.

[3] Chase TN, “Levodopa therapy: consequences of the nonphysiologic replacement of dopamine”, Neurology, 50(Suppl5): S17-S25, 1998

[4] Fahn S., Elton R.L., “Unified Parkinson’s Disease Rating Scale”. In Fahn S (Ed), Recent Developments in Parkinson’s Disease, MacMillan Healthcare Information, 153-163, 1987

[5] Jolliffe, I.T., Principal Components Analysis. New York: Springer-Verlag, 1986.

[6] Lang AE, Lozano AM, “Parkinson's disease. First of two parts”, N Engl J Med, 339(16): 1044-1053, 1998

[7] Lang AE, Lozano AM, “Parkinson's disease. Second of two parts”, N Engl J Med, 339(16): 1130-1143, 1998

[8] Mouradian MM, Heuser IJ, Baronti F, Chase TN, “Modification of central dopaminergic mechanisms by continuous levodopa therapy for advanced Parkinson's disease”, Ann Neurol, 27(1): 18-23, 1990.

[9] Obeso JA, Olanow CW, Nutt JG, “Levodopa motor complications in Parkinson's disease”, Trends Neurosci, 23: 2-7, 2000

[10] Parkinson Study Group, “Evaluation of dyskinesias in a pilot, randomized, placebo-controlled trial of remacemide in advanced Parkinson’s disease”, Arch Neurol, 58(10): 1660-1668, 2001.

[11] Sammon MP, Bruce EN, “Vagal afferent activity increases dynamical dimension of respiration in rats”, J Appl Physiol, 70(4): 1748-1762, 1991.

[12] Sekine M, Tamura T, Akay M, Fujimoto T, Togawa T, Fukui Y, “Discrimination of walking patterns using wavelet-based fractal analysis”, IEEE Trans on Neural Systems and Rehabilitation Engineering

[13] Standaert DG, Young AB, in Goodman and Gilman's Pharmacological Basis of Therapeutics, Tenth Edition, Hardman JG and Limbird LE, Editors, McGraw-Hill, 549-620, 2001

Sammon’s map of Bradykinesia-related features

Time (min)

BradykinesiaScore

0 9030 180150120600123

LevodopaIntake

OFF

WEARING-OFFON

OFFMotor FluctuationCycle

Sammon’s map of Bradykinesia-related features

Time (min)

BradykinesiaScore

0 9030 180150120600123

LevodopaIntake

OFF

WEARING-OFFON

OFFMotor FluctuationCycle

Time (min)

BradykinesiaScore

0 9030 180150120600123

0 9030 180150120600 9030 180150120600123

LevodopaIntake

OFF

WEARING-OFFON

OFF

LevodopaIntake

LevodopaIntake

OFF

WEARING-OFFON

OFFOFF

WEARING-OFFON

OFFMotor FluctuationCycle

Figure 3 Two-dimensional projection of bradykinesia-relatedfeatures for a representative subject. Axis units are unspecifiedbecause they result from abstract transformations of normalized dataand hence lack a direct physical interpretation. Scatter plot is color-coded according to clinical score assigned for each testing interval, asshown in the timeline (0=normal, 1=mild, 2=moderate, 3=severe). Aschematic of the motor fluctuation cycle is shown for reference.